program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.11.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios19>(tensor<fp16, [1, 1024]> attention_mask, tensor<int32, [1, 1024]> input_ids) {
            tensor<fp16, [32768, 128]> encoder_sin_cached = const()[name = string("encoder_sin_cached"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [32768, 128]> encoder_cos_cached = const()[name = string("encoder_cos_cached"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8388736)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_0_self_attn_q_proj_weight = const()[name = string("encoder_layers_0_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16777408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_0_self_attn_k_proj_weight = const()[name = string("encoder_layers_0_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20971776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_0_self_attn_v_proj_weight = const()[name = string("encoder_layers_0_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23068992)))];
            tensor<fp16, [128]> encoder_layers_0_self_attn_q_norm_weight = const()[name = string("encoder_layers_0_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25166208)))];
            tensor<fp16, [128]> encoder_layers_0_self_attn_k_norm_weight = const()[name = string("encoder_layers_0_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25166528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_0_mlp_gate_proj_weight = const()[name = string("encoder_layers_0_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25166848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_0_mlp_up_proj_weight = const()[name = string("encoder_layers_0_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31458368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_0_mlp_down_proj_weight = const()[name = string("encoder_layers_0_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37749888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_1_self_attn_q_proj_weight = const()[name = string("encoder_layers_1_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44041408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_1_self_attn_k_proj_weight = const()[name = string("encoder_layers_1_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48235776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_1_self_attn_v_proj_weight = const()[name = string("encoder_layers_1_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50332992)))];
            tensor<fp16, [128]> encoder_layers_1_self_attn_q_norm_weight = const()[name = string("encoder_layers_1_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52430208)))];
            tensor<fp16, [128]> encoder_layers_1_self_attn_k_norm_weight = const()[name = string("encoder_layers_1_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52430528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_1_mlp_gate_proj_weight = const()[name = string("encoder_layers_1_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52430848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_1_mlp_up_proj_weight = const()[name = string("encoder_layers_1_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58722368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_1_mlp_down_proj_weight = const()[name = string("encoder_layers_1_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65013888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_2_self_attn_q_proj_weight = const()[name = string("encoder_layers_2_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71305408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_2_self_attn_k_proj_weight = const()[name = string("encoder_layers_2_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75499776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_2_self_attn_v_proj_weight = const()[name = string("encoder_layers_2_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77596992)))];
            tensor<fp16, [128]> encoder_layers_2_self_attn_q_norm_weight = const()[name = string("encoder_layers_2_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79694208)))];
            tensor<fp16, [128]> encoder_layers_2_self_attn_k_norm_weight = const()[name = string("encoder_layers_2_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79694528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_2_mlp_gate_proj_weight = const()[name = string("encoder_layers_2_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79694848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_2_mlp_up_proj_weight = const()[name = string("encoder_layers_2_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85986368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_2_mlp_down_proj_weight = const()[name = string("encoder_layers_2_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92277888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_3_self_attn_q_proj_weight = const()[name = string("encoder_layers_3_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98569408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_3_self_attn_k_proj_weight = const()[name = string("encoder_layers_3_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102763776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_3_self_attn_v_proj_weight = const()[name = string("encoder_layers_3_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104860992)))];
            tensor<fp16, [128]> encoder_layers_3_self_attn_q_norm_weight = const()[name = string("encoder_layers_3_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106958208)))];
            tensor<fp16, [128]> encoder_layers_3_self_attn_k_norm_weight = const()[name = string("encoder_layers_3_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106958528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_3_mlp_gate_proj_weight = const()[name = string("encoder_layers_3_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106958848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_3_mlp_up_proj_weight = const()[name = string("encoder_layers_3_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113250368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_3_mlp_down_proj_weight = const()[name = string("encoder_layers_3_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119541888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_4_self_attn_q_proj_weight = const()[name = string("encoder_layers_4_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125833408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_4_self_attn_k_proj_weight = const()[name = string("encoder_layers_4_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130027776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_4_self_attn_v_proj_weight = const()[name = string("encoder_layers_4_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132124992)))];
            tensor<fp16, [128]> encoder_layers_4_self_attn_q_norm_weight = const()[name = string("encoder_layers_4_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134222208)))];
            tensor<fp16, [128]> encoder_layers_4_self_attn_k_norm_weight = const()[name = string("encoder_layers_4_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134222528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_4_mlp_gate_proj_weight = const()[name = string("encoder_layers_4_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134222848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_4_mlp_up_proj_weight = const()[name = string("encoder_layers_4_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140514368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_4_mlp_down_proj_weight = const()[name = string("encoder_layers_4_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146805888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_5_self_attn_q_proj_weight = const()[name = string("encoder_layers_5_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153097408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_5_self_attn_k_proj_weight = const()[name = string("encoder_layers_5_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157291776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_5_self_attn_v_proj_weight = const()[name = string("encoder_layers_5_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159388992)))];
            tensor<fp16, [128]> encoder_layers_5_self_attn_q_norm_weight = const()[name = string("encoder_layers_5_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161486208)))];
            tensor<fp16, [128]> encoder_layers_5_self_attn_k_norm_weight = const()[name = string("encoder_layers_5_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161486528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_5_mlp_gate_proj_weight = const()[name = string("encoder_layers_5_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161486848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_5_mlp_up_proj_weight = const()[name = string("encoder_layers_5_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167778368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_5_mlp_down_proj_weight = const()[name = string("encoder_layers_5_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174069888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_6_self_attn_q_proj_weight = const()[name = string("encoder_layers_6_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180361408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_6_self_attn_k_proj_weight = const()[name = string("encoder_layers_6_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184555776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_6_self_attn_v_proj_weight = const()[name = string("encoder_layers_6_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186652992)))];
            tensor<fp16, [128]> encoder_layers_6_self_attn_q_norm_weight = const()[name = string("encoder_layers_6_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188750208)))];
            tensor<fp16, [128]> encoder_layers_6_self_attn_k_norm_weight = const()[name = string("encoder_layers_6_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188750528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_6_mlp_gate_proj_weight = const()[name = string("encoder_layers_6_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188750848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_6_mlp_up_proj_weight = const()[name = string("encoder_layers_6_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195042368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_6_mlp_down_proj_weight = const()[name = string("encoder_layers_6_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201333888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_7_self_attn_q_proj_weight = const()[name = string("encoder_layers_7_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207625408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_7_self_attn_k_proj_weight = const()[name = string("encoder_layers_7_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211819776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_7_self_attn_v_proj_weight = const()[name = string("encoder_layers_7_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213916992)))];
            tensor<fp16, [128]> encoder_layers_7_self_attn_q_norm_weight = const()[name = string("encoder_layers_7_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216014208)))];
            tensor<fp16, [128]> encoder_layers_7_self_attn_k_norm_weight = const()[name = string("encoder_layers_7_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216014528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_7_mlp_gate_proj_weight = const()[name = string("encoder_layers_7_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216014848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_7_mlp_up_proj_weight = const()[name = string("encoder_layers_7_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222306368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_7_mlp_down_proj_weight = const()[name = string("encoder_layers_7_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228597888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_8_self_attn_q_proj_weight = const()[name = string("encoder_layers_8_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234889408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_8_self_attn_k_proj_weight = const()[name = string("encoder_layers_8_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239083776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_8_self_attn_v_proj_weight = const()[name = string("encoder_layers_8_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241180992)))];
            tensor<fp16, [128]> encoder_layers_8_self_attn_q_norm_weight = const()[name = string("encoder_layers_8_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243278208)))];
            tensor<fp16, [128]> encoder_layers_8_self_attn_k_norm_weight = const()[name = string("encoder_layers_8_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243278528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_8_mlp_gate_proj_weight = const()[name = string("encoder_layers_8_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243278848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_8_mlp_up_proj_weight = const()[name = string("encoder_layers_8_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249570368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_8_mlp_down_proj_weight = const()[name = string("encoder_layers_8_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255861888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_9_self_attn_q_proj_weight = const()[name = string("encoder_layers_9_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262153408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_9_self_attn_k_proj_weight = const()[name = string("encoder_layers_9_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266347776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_9_self_attn_v_proj_weight = const()[name = string("encoder_layers_9_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268444992)))];
            tensor<fp16, [128]> encoder_layers_9_self_attn_q_norm_weight = const()[name = string("encoder_layers_9_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270542208)))];
            tensor<fp16, [128]> encoder_layers_9_self_attn_k_norm_weight = const()[name = string("encoder_layers_9_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270542528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_9_mlp_gate_proj_weight = const()[name = string("encoder_layers_9_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270542848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_9_mlp_up_proj_weight = const()[name = string("encoder_layers_9_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276834368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_9_mlp_down_proj_weight = const()[name = string("encoder_layers_9_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283125888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_10_self_attn_q_proj_weight = const()[name = string("encoder_layers_10_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289417408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_10_self_attn_k_proj_weight = const()[name = string("encoder_layers_10_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293611776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_10_self_attn_v_proj_weight = const()[name = string("encoder_layers_10_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295708992)))];
            tensor<fp16, [128]> encoder_layers_10_self_attn_q_norm_weight = const()[name = string("encoder_layers_10_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297806208)))];
            tensor<fp16, [128]> encoder_layers_10_self_attn_k_norm_weight = const()[name = string("encoder_layers_10_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297806528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_10_mlp_gate_proj_weight = const()[name = string("encoder_layers_10_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297806848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_10_mlp_up_proj_weight = const()[name = string("encoder_layers_10_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304098368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_10_mlp_down_proj_weight = const()[name = string("encoder_layers_10_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310389888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_11_self_attn_q_proj_weight = const()[name = string("encoder_layers_11_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316681408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_11_self_attn_k_proj_weight = const()[name = string("encoder_layers_11_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320875776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_11_self_attn_v_proj_weight = const()[name = string("encoder_layers_11_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322972992)))];
            tensor<fp16, [128]> encoder_layers_11_self_attn_q_norm_weight = const()[name = string("encoder_layers_11_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325070208)))];
            tensor<fp16, [128]> encoder_layers_11_self_attn_k_norm_weight = const()[name = string("encoder_layers_11_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325070528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_11_mlp_gate_proj_weight = const()[name = string("encoder_layers_11_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325070848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_11_mlp_up_proj_weight = const()[name = string("encoder_layers_11_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331362368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_11_mlp_down_proj_weight = const()[name = string("encoder_layers_11_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337653888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_12_self_attn_q_proj_weight = const()[name = string("encoder_layers_12_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343945408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_12_self_attn_k_proj_weight = const()[name = string("encoder_layers_12_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348139776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_12_self_attn_v_proj_weight = const()[name = string("encoder_layers_12_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350236992)))];
            tensor<fp16, [128]> encoder_layers_12_self_attn_q_norm_weight = const()[name = string("encoder_layers_12_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352334208)))];
            tensor<fp16, [128]> encoder_layers_12_self_attn_k_norm_weight = const()[name = string("encoder_layers_12_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352334528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_12_mlp_gate_proj_weight = const()[name = string("encoder_layers_12_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352334848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_12_mlp_up_proj_weight = const()[name = string("encoder_layers_12_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358626368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_12_mlp_down_proj_weight = const()[name = string("encoder_layers_12_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364917888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_13_self_attn_q_proj_weight = const()[name = string("encoder_layers_13_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371209408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_13_self_attn_k_proj_weight = const()[name = string("encoder_layers_13_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375403776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_13_self_attn_v_proj_weight = const()[name = string("encoder_layers_13_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377500992)))];
            tensor<fp16, [128]> encoder_layers_13_self_attn_q_norm_weight = const()[name = string("encoder_layers_13_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379598208)))];
            tensor<fp16, [128]> encoder_layers_13_self_attn_k_norm_weight = const()[name = string("encoder_layers_13_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379598528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_13_mlp_gate_proj_weight = const()[name = string("encoder_layers_13_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379598848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_13_mlp_up_proj_weight = const()[name = string("encoder_layers_13_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385890368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_13_mlp_down_proj_weight = const()[name = string("encoder_layers_13_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392181888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_14_self_attn_q_proj_weight = const()[name = string("encoder_layers_14_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398473408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_14_self_attn_k_proj_weight = const()[name = string("encoder_layers_14_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402667776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_14_self_attn_v_proj_weight = const()[name = string("encoder_layers_14_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404764992)))];
            tensor<fp16, [128]> encoder_layers_14_self_attn_q_norm_weight = const()[name = string("encoder_layers_14_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406862208)))];
            tensor<fp16, [128]> encoder_layers_14_self_attn_k_norm_weight = const()[name = string("encoder_layers_14_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406862528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_14_mlp_gate_proj_weight = const()[name = string("encoder_layers_14_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406862848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_14_mlp_up_proj_weight = const()[name = string("encoder_layers_14_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413154368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_14_mlp_down_proj_weight = const()[name = string("encoder_layers_14_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419445888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_15_self_attn_q_proj_weight = const()[name = string("encoder_layers_15_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425737408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_15_self_attn_k_proj_weight = const()[name = string("encoder_layers_15_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429931776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_15_self_attn_v_proj_weight = const()[name = string("encoder_layers_15_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432028992)))];
            tensor<fp16, [128]> encoder_layers_15_self_attn_q_norm_weight = const()[name = string("encoder_layers_15_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434126208)))];
            tensor<fp16, [128]> encoder_layers_15_self_attn_k_norm_weight = const()[name = string("encoder_layers_15_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434126528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_15_mlp_gate_proj_weight = const()[name = string("encoder_layers_15_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434126848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_15_mlp_up_proj_weight = const()[name = string("encoder_layers_15_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440418368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_15_mlp_down_proj_weight = const()[name = string("encoder_layers_15_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446709888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_16_self_attn_q_proj_weight = const()[name = string("encoder_layers_16_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453001408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_16_self_attn_k_proj_weight = const()[name = string("encoder_layers_16_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457195776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_16_self_attn_v_proj_weight = const()[name = string("encoder_layers_16_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459292992)))];
            tensor<fp16, [128]> encoder_layers_16_self_attn_q_norm_weight = const()[name = string("encoder_layers_16_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461390208)))];
            tensor<fp16, [128]> encoder_layers_16_self_attn_k_norm_weight = const()[name = string("encoder_layers_16_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461390528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_16_mlp_gate_proj_weight = const()[name = string("encoder_layers_16_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461390848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_16_mlp_up_proj_weight = const()[name = string("encoder_layers_16_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467682368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_16_mlp_down_proj_weight = const()[name = string("encoder_layers_16_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473973888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_17_self_attn_q_proj_weight = const()[name = string("encoder_layers_17_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480265408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_17_self_attn_k_proj_weight = const()[name = string("encoder_layers_17_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484459776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_17_self_attn_v_proj_weight = const()[name = string("encoder_layers_17_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486556992)))];
            tensor<fp16, [128]> encoder_layers_17_self_attn_q_norm_weight = const()[name = string("encoder_layers_17_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488654208)))];
            tensor<fp16, [128]> encoder_layers_17_self_attn_k_norm_weight = const()[name = string("encoder_layers_17_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488654528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_17_mlp_gate_proj_weight = const()[name = string("encoder_layers_17_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488654848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_17_mlp_up_proj_weight = const()[name = string("encoder_layers_17_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494946368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_17_mlp_down_proj_weight = const()[name = string("encoder_layers_17_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501237888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_18_self_attn_q_proj_weight = const()[name = string("encoder_layers_18_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(507529408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_18_self_attn_k_proj_weight = const()[name = string("encoder_layers_18_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511723776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_18_self_attn_v_proj_weight = const()[name = string("encoder_layers_18_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513820992)))];
            tensor<fp16, [128]> encoder_layers_18_self_attn_q_norm_weight = const()[name = string("encoder_layers_18_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515918208)))];
            tensor<fp16, [128]> encoder_layers_18_self_attn_k_norm_weight = const()[name = string("encoder_layers_18_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515918528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_18_mlp_gate_proj_weight = const()[name = string("encoder_layers_18_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515918848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_18_mlp_up_proj_weight = const()[name = string("encoder_layers_18_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522210368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_18_mlp_down_proj_weight = const()[name = string("encoder_layers_18_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528501888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_19_self_attn_q_proj_weight = const()[name = string("encoder_layers_19_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534793408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_19_self_attn_k_proj_weight = const()[name = string("encoder_layers_19_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538987776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_19_self_attn_v_proj_weight = const()[name = string("encoder_layers_19_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541084992)))];
            tensor<fp16, [128]> encoder_layers_19_self_attn_q_norm_weight = const()[name = string("encoder_layers_19_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543182208)))];
            tensor<fp16, [128]> encoder_layers_19_self_attn_k_norm_weight = const()[name = string("encoder_layers_19_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543182528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_19_mlp_gate_proj_weight = const()[name = string("encoder_layers_19_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543182848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_19_mlp_up_proj_weight = const()[name = string("encoder_layers_19_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549474368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_19_mlp_down_proj_weight = const()[name = string("encoder_layers_19_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555765888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_20_self_attn_q_proj_weight = const()[name = string("encoder_layers_20_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562057408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_20_self_attn_k_proj_weight = const()[name = string("encoder_layers_20_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566251776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_20_self_attn_v_proj_weight = const()[name = string("encoder_layers_20_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568348992)))];
            tensor<fp16, [128]> encoder_layers_20_self_attn_q_norm_weight = const()[name = string("encoder_layers_20_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570446208)))];
            tensor<fp16, [128]> encoder_layers_20_self_attn_k_norm_weight = const()[name = string("encoder_layers_20_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570446528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_20_mlp_gate_proj_weight = const()[name = string("encoder_layers_20_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570446848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_20_mlp_up_proj_weight = const()[name = string("encoder_layers_20_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576738368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_20_mlp_down_proj_weight = const()[name = string("encoder_layers_20_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583029888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_21_self_attn_q_proj_weight = const()[name = string("encoder_layers_21_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589321408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_21_self_attn_k_proj_weight = const()[name = string("encoder_layers_21_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593515776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_21_self_attn_v_proj_weight = const()[name = string("encoder_layers_21_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595612992)))];
            tensor<fp16, [128]> encoder_layers_21_self_attn_q_norm_weight = const()[name = string("encoder_layers_21_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597710208)))];
            tensor<fp16, [128]> encoder_layers_21_self_attn_k_norm_weight = const()[name = string("encoder_layers_21_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597710528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_21_mlp_gate_proj_weight = const()[name = string("encoder_layers_21_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597710848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_21_mlp_up_proj_weight = const()[name = string("encoder_layers_21_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604002368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_21_mlp_down_proj_weight = const()[name = string("encoder_layers_21_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610293888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_22_self_attn_q_proj_weight = const()[name = string("encoder_layers_22_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616585408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_22_self_attn_k_proj_weight = const()[name = string("encoder_layers_22_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620779776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_22_self_attn_v_proj_weight = const()[name = string("encoder_layers_22_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(622876992)))];
            tensor<fp16, [128]> encoder_layers_22_self_attn_q_norm_weight = const()[name = string("encoder_layers_22_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624974208)))];
            tensor<fp16, [128]> encoder_layers_22_self_attn_k_norm_weight = const()[name = string("encoder_layers_22_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624974528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_22_mlp_gate_proj_weight = const()[name = string("encoder_layers_22_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624974848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_22_mlp_up_proj_weight = const()[name = string("encoder_layers_22_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631266368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_22_mlp_down_proj_weight = const()[name = string("encoder_layers_22_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637557888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_23_self_attn_q_proj_weight = const()[name = string("encoder_layers_23_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643849408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_23_self_attn_k_proj_weight = const()[name = string("encoder_layers_23_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648043776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_23_self_attn_v_proj_weight = const()[name = string("encoder_layers_23_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650140992)))];
            tensor<fp16, [128]> encoder_layers_23_self_attn_q_norm_weight = const()[name = string("encoder_layers_23_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652238208)))];
            tensor<fp16, [128]> encoder_layers_23_self_attn_k_norm_weight = const()[name = string("encoder_layers_23_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652238528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_23_mlp_gate_proj_weight = const()[name = string("encoder_layers_23_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652238848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_23_mlp_up_proj_weight = const()[name = string("encoder_layers_23_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658530368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_23_mlp_down_proj_weight = const()[name = string("encoder_layers_23_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664821888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_24_self_attn_q_proj_weight = const()[name = string("encoder_layers_24_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(671113408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_24_self_attn_k_proj_weight = const()[name = string("encoder_layers_24_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675307776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_24_self_attn_v_proj_weight = const()[name = string("encoder_layers_24_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677404992)))];
            tensor<fp16, [128]> encoder_layers_24_self_attn_q_norm_weight = const()[name = string("encoder_layers_24_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679502208)))];
            tensor<fp16, [128]> encoder_layers_24_self_attn_k_norm_weight = const()[name = string("encoder_layers_24_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679502528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_24_mlp_gate_proj_weight = const()[name = string("encoder_layers_24_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679502848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_24_mlp_up_proj_weight = const()[name = string("encoder_layers_24_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685794368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_24_mlp_down_proj_weight = const()[name = string("encoder_layers_24_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692085888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_25_self_attn_q_proj_weight = const()[name = string("encoder_layers_25_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698377408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_25_self_attn_k_proj_weight = const()[name = string("encoder_layers_25_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702571776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_25_self_attn_v_proj_weight = const()[name = string("encoder_layers_25_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704668992)))];
            tensor<fp16, [128]> encoder_layers_25_self_attn_q_norm_weight = const()[name = string("encoder_layers_25_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706766208)))];
            tensor<fp16, [128]> encoder_layers_25_self_attn_k_norm_weight = const()[name = string("encoder_layers_25_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706766528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_25_mlp_gate_proj_weight = const()[name = string("encoder_layers_25_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706766848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_25_mlp_up_proj_weight = const()[name = string("encoder_layers_25_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(713058368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_25_mlp_down_proj_weight = const()[name = string("encoder_layers_25_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719349888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_26_self_attn_q_proj_weight = const()[name = string("encoder_layers_26_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(725641408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_26_self_attn_k_proj_weight = const()[name = string("encoder_layers_26_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729835776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_26_self_attn_v_proj_weight = const()[name = string("encoder_layers_26_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(731932992)))];
            tensor<fp16, [128]> encoder_layers_26_self_attn_q_norm_weight = const()[name = string("encoder_layers_26_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(734030208)))];
            tensor<fp16, [128]> encoder_layers_26_self_attn_k_norm_weight = const()[name = string("encoder_layers_26_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(734030528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_26_mlp_gate_proj_weight = const()[name = string("encoder_layers_26_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(734030848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_26_mlp_up_proj_weight = const()[name = string("encoder_layers_26_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740322368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_26_mlp_down_proj_weight = const()[name = string("encoder_layers_26_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746613888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_27_self_attn_q_proj_weight = const()[name = string("encoder_layers_27_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752905408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_27_self_attn_k_proj_weight = const()[name = string("encoder_layers_27_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757099776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_27_self_attn_v_proj_weight = const()[name = string("encoder_layers_27_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(759196992)))];
            tensor<fp16, [128]> encoder_layers_27_self_attn_q_norm_weight = const()[name = string("encoder_layers_27_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(761294208)))];
            tensor<fp16, [128]> encoder_layers_27_self_attn_k_norm_weight = const()[name = string("encoder_layers_27_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(761294528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_27_mlp_gate_proj_weight = const()[name = string("encoder_layers_27_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(761294848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_27_mlp_up_proj_weight = const()[name = string("encoder_layers_27_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(767586368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_27_mlp_down_proj_weight = const()[name = string("encoder_layers_27_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773877888)))];
            int32 var_17 = const()[name = string("op_17"), val = int32(-1)];
            int32 var_19 = const()[name = string("op_19"), val = int32(1)];
            int32 var_82_batch_dims_0 = const()[name = string("op_82_batch_dims_0"), val = int32(0)];
            bool var_82_validate_indices_0 = const()[name = string("op_82_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 1024]> encoder_embed_tokens_weight_to_fp16 = const()[name = string("encoder_embed_tokens_weight_to_fp16"), val = tensor<fp16, [151936, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(780169408)))];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 1024]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 1024]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 1024]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 var_82_cast_fp16_axis_0 = const()[name = string("op_82_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 1024, 1024]> var_82_cast_fp16 = gather(axis = var_82_cast_fp16_axis_0, batch_dims = var_82_batch_dims_0, indices = select_0, validate_indices = var_82_validate_indices_0, x = encoder_embed_tokens_weight_to_fp16)[name = string("op_82_cast_fp16")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 1024]> fill_like_0_cast_fp16 = fill_like(ref_tensor = attention_mask, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            bool var_85_exclusive_0 = const()[name = string("op_85_exclusive_0"), val = bool(false)];
            bool var_85_reverse_0 = const()[name = string("op_85_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 1024]> var_85_cast_fp16 = cumsum(axis = var_19, exclusive = var_85_exclusive_0, reverse = var_85_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_85_cast_fp16")];
            fp16 var_86_to_fp16 = const()[name = string("op_86_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 1024]> var_87_cast_fp16 = sub(x = var_85_cast_fp16, y = var_86_to_fp16)[name = string("op_87_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            tensor<int32, [2]> pos_begin_0 = const()[name = string("pos_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [2]> pos_end_0 = const()[name = string("pos_end_0"), val = tensor<int32, [2]>([1, 1024])];
            tensor<bool, [2]> pos_end_mask_0 = const()[name = string("pos_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<bool, [2]> pos_squeeze_mask_0 = const()[name = string("pos_squeeze_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<int32, [1, 1024]> var_87_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = var_87_cast_fp16)[name = string("cast_232")];
            tensor<int32, [1024]> pos = slice_by_index(begin = pos_begin_0, end = pos_end_0, end_mask = pos_end_mask_0, squeeze_mask = pos_squeeze_mask_0, x = var_87_cast_fp16_to_int32)[name = string("pos")];
            int32 var_90_batch_dims_0 = const()[name = string("op_90_batch_dims_0"), val = int32(0)];
            bool var_90_validate_indices_0 = const()[name = string("op_90_validate_indices_0"), val = bool(false)];
            int32 greater_equal_1_y_0 = const()[name = string("greater_equal_1_y_0"), val = int32(0)];
            tensor<bool, [1024]> greater_equal_1 = greater_equal(x = pos, y = greater_equal_1_y_0)[name = string("greater_equal_1")];
            int32 slice_by_index_1 = const()[name = string("slice_by_index_1"), val = int32(32768)];
            tensor<int32, [1024]> add_1 = add(x = pos, y = slice_by_index_1)[name = string("add_1")];
            tensor<int32, [1024]> select_1 = select(a = pos, b = add_1, cond = greater_equal_1)[name = string("select_1")];
            int32 var_90_axis_0 = const()[name = string("op_90_axis_0"), val = int32(0)];
            tensor<fp16, [1024, 128]> var_90 = gather(axis = var_90_axis_0, batch_dims = var_90_batch_dims_0, indices = select_1, validate_indices = var_90_validate_indices_0, x = encoder_cos_cached)[name = string("op_90")];
            tensor<int32, [4]> var_91 = const()[name = string("op_91"), val = tensor<int32, [4]>([1, 1, 1024, 128])];
            tensor<fp16, [1, 1, 1024, 128]> cos = reshape(shape = var_91, x = var_90)[name = string("cos")];
            int32 var_93_batch_dims_0 = const()[name = string("op_93_batch_dims_0"), val = int32(0)];
            bool var_93_validate_indices_0 = const()[name = string("op_93_validate_indices_0"), val = bool(false)];
            int32 var_93_axis_0 = const()[name = string("op_93_axis_0"), val = int32(0)];
            tensor<fp16, [1024, 128]> var_93 = gather(axis = var_93_axis_0, batch_dims = var_93_batch_dims_0, indices = select_1, validate_indices = var_93_validate_indices_0, x = encoder_sin_cached)[name = string("op_93")];
            tensor<int32, [4]> var_94 = const()[name = string("op_94"), val = tensor<int32, [4]>([1, 1, 1024, 128])];
            tensor<fp16, [1, 1, 1024, 128]> sin = reshape(shape = var_94, x = var_93)[name = string("sin")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 1024]> var_97_cast_fp16 = sub(x = var_11_to_fp16, y = attention_mask)[name = string("op_97_cast_fp16")];
            fp16 var_99_to_fp16 = const()[name = string("op_99_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 1024]> key_pad_cast_fp16 = mul(x = var_97_cast_fp16, y = var_99_to_fp16)[name = string("key_pad_cast_fp16")];
            tensor<int32, [4]> var_101 = const()[name = string("op_101"), val = tensor<int32, [4]>([1, 1, 1, 1024])];
            tensor<fp16, [1, 1, 1, 1024]> var_102_cast_fp16 = reshape(shape = var_101, x = key_pad_cast_fp16)[name = string("op_102_cast_fp16")];
            tensor<int32, [4]> causal_mask_reps_0 = const()[name = string("causal_mask_reps_0"), val = tensor<int32, [4]>([1, 1, 1024, 1])];
            tensor<fp16, [1, 1, 1024, 1024]> causal_mask_cast_fp16 = tile(reps = causal_mask_reps_0, x = var_102_cast_fp16)[name = string("causal_mask_cast_fp16")];
            fp16 var_5_promoted_to_fp16 = const()[name = string("op_5_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_128_cast_fp16 = pow(x = var_82_cast_fp16, y = var_5_promoted_to_fp16)[name = string("op_128_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_128_cast_fp16)[name = string("var_1_cast_fp16")];
            fp16 var_131_to_fp16 = const()[name = string("op_131_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_132_cast_fp16 = add(x = var_1_cast_fp16, y = var_131_to_fp16)[name = string("op_132_cast_fp16")];
            fp32 var_133_epsilon_0 = const()[name = string("op_133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_133_cast_fp16 = rsqrt(epsilon = var_133_epsilon_0, x = var_132_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_3_cast_fp16 = mul(x = var_82_cast_fp16, y = var_133_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_0_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_0_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1091334400)))];
            tensor<fp16, [1, 1024, 1024]> var_136_cast_fp16 = mul(x = x_3_cast_fp16, y = encoder_layers_0_input_layernorm_weight_promoted_to_fp16)[name = string("op_136_cast_fp16")];
            tensor<int32, [3]> var_141 = const()[name = string("op_141"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1_axes_0 = const()[name = string("input_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_142 = transpose(perm = var_141, x = var_136_cast_fp16)[name = string("transpose_251")];
            tensor<fp16, [1, 1024, 1, 1024]> input_1 = expand_dims(axes = input_1_axes_0, x = var_142)[name = string("input_1")];
            string var_149_pad_type_0 = const()[name = string("op_149_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_149_strides_0 = const()[name = string("op_149_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_149_pad_0 = const()[name = string("op_149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_149_dilations_0 = const()[name = string("op_149_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_149_groups_0 = const()[name = string("op_149_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_149 = conv(dilations = var_149_dilations_0, groups = var_149_groups_0, pad = var_149_pad_0, pad_type = var_149_pad_type_0, strides = var_149_strides_0, weight = encoder_layers_0_self_attn_q_proj_weight, x = input_1)[name = string("op_149")];
            tensor<int32, [4]> var_150 = const()[name = string("op_150"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_151 = reshape(shape = var_150, x = var_149)[name = string("op_151")];
            tensor<int32, [4]> var_152 = const()[name = string("op_152"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_159_pad_type_0 = const()[name = string("op_159_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_159_strides_0 = const()[name = string("op_159_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_159_pad_0 = const()[name = string("op_159_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_159_dilations_0 = const()[name = string("op_159_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_159_groups_0 = const()[name = string("op_159_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_159 = conv(dilations = var_159_dilations_0, groups = var_159_groups_0, pad = var_159_pad_0, pad_type = var_159_pad_type_0, strides = var_159_strides_0, weight = encoder_layers_0_self_attn_k_proj_weight, x = input_1)[name = string("op_159")];
            tensor<int32, [4]> var_160 = const()[name = string("op_160"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_161 = reshape(shape = var_160, x = var_159)[name = string("op_161")];
            tensor<int32, [4]> var_162 = const()[name = string("op_162"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_169_pad_type_0 = const()[name = string("op_169_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_169_strides_0 = const()[name = string("op_169_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_169_pad_0 = const()[name = string("op_169_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_169_dilations_0 = const()[name = string("op_169_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_169_groups_0 = const()[name = string("op_169_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_169 = conv(dilations = var_169_dilations_0, groups = var_169_groups_0, pad = var_169_pad_0, pad_type = var_169_pad_type_0, strides = var_169_strides_0, weight = encoder_layers_0_self_attn_v_proj_weight, x = input_1)[name = string("op_169")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_171 = reshape(shape = var_170, x = var_169)[name = string("op_171")];
            tensor<int32, [4]> var_172 = const()[name = string("op_172"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_1_to_fp16 = const()[name = string("op_5_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_1 = transpose(perm = var_152, x = var_151)[name = string("transpose_250")];
            tensor<fp16, [1, 16, 1024, 128]> var_178_cast_fp16 = pow(x = q_1, y = var_5_promoted_1_to_fp16)[name = string("op_178_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_178_cast_fp16)[name = string("var_3_cast_fp16")];
            fp16 var_181_to_fp16 = const()[name = string("op_181_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_182_cast_fp16 = add(x = var_3_cast_fp16, y = var_181_to_fp16)[name = string("op_182_cast_fp16")];
            fp32 var_183_epsilon_0 = const()[name = string("op_183_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_183_cast_fp16 = rsqrt(epsilon = var_183_epsilon_0, x = var_182_cast_fp16)[name = string("op_183_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_11_cast_fp16 = mul(x = q_1, y = var_183_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_3 = mul(x = x_11_cast_fp16, y = encoder_layers_0_self_attn_q_norm_weight)[name = string("q_3")];
            fp16 var_5_promoted_2_to_fp16 = const()[name = string("op_5_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_1 = transpose(perm = var_162, x = var_161)[name = string("transpose_249")];
            tensor<fp16, [1, 8, 1024, 128]> var_191_cast_fp16 = pow(x = k_1, y = var_5_promoted_2_to_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_191_cast_fp16)[name = string("var_5_cast_fp16")];
            fp16 var_194_to_fp16 = const()[name = string("op_194_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_195_cast_fp16 = add(x = var_5_cast_fp16, y = var_194_to_fp16)[name = string("op_195_cast_fp16")];
            fp32 var_196_epsilon_0 = const()[name = string("op_196_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_196_cast_fp16 = rsqrt(epsilon = var_196_epsilon_0, x = var_195_cast_fp16)[name = string("op_196_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_17_cast_fp16 = mul(x = k_1, y = var_196_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_3 = mul(x = x_17_cast_fp16, y = encoder_layers_0_self_attn_k_norm_weight)[name = string("k_3")];
            tensor<fp16, [1, 16, 1024, 128]> var_200 = mul(x = q_3, y = cos)[name = string("op_200")];
            tensor<int32, [2]> var_201_split_sizes_0 = const()[name = string("op_201_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_201_axis_0 = const()[name = string("op_201_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_201_0, tensor<fp16, [1, 16, 1024, 64]> var_201_1 = split(axis = var_201_axis_0, split_sizes = var_201_split_sizes_0, x = q_3)[name = string("op_201")];
            fp16 const_3_promoted = const()[name = string("const_3_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_203 = mul(x = var_201_1, y = const_3_promoted)[name = string("op_203")];
            bool var_205_interleave_0 = const()[name = string("op_205_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_205 = concat(axis = var_17, interleave = var_205_interleave_0, values = (var_203, var_201_0))[name = string("op_205")];
            tensor<fp16, [1, 16, 1024, 128]> var_206 = mul(x = var_205, y = sin)[name = string("op_206")];
            tensor<fp16, [1, 16, 1024, 128]> query_1 = add(x = var_200, y = var_206)[name = string("query_1")];
            tensor<fp16, [1, 8, 1024, 128]> var_208 = mul(x = k_3, y = cos)[name = string("op_208")];
            tensor<int32, [2]> var_209_split_sizes_0 = const()[name = string("op_209_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_209_axis_0 = const()[name = string("op_209_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_209_0, tensor<fp16, [1, 8, 1024, 64]> var_209_1 = split(axis = var_209_axis_0, split_sizes = var_209_split_sizes_0, x = k_3)[name = string("op_209")];
            fp16 const_4_promoted = const()[name = string("const_4_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_211 = mul(x = var_209_1, y = const_4_promoted)[name = string("op_211")];
            bool var_213_interleave_0 = const()[name = string("op_213_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_213 = concat(axis = var_17, interleave = var_213_interleave_0, values = (var_211, var_209_0))[name = string("op_213")];
            tensor<fp16, [1, 8, 1024, 128]> var_214 = mul(x = var_213, y = sin)[name = string("op_214")];
            tensor<fp16, [1, 8, 1024, 128]> x_19 = add(x = var_208, y = var_214)[name = string("x_19")];
            tensor<int32, [1]> var_216_axes_0 = const()[name = string("op_216_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_216 = expand_dims(axes = var_216_axes_0, x = x_19)[name = string("op_216")];
            tensor<int32, [5]> x_21_reps_0 = const()[name = string("x_21_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_21 = tile(reps = x_21_reps_0, x = var_216)[name = string("x_21")];
            tensor<int32, [4]> var_219 = const()[name = string("op_219"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_1 = reshape(shape = var_219, x = x_21)[name = string("key_1")];
            tensor<int32, [1]> var_221_axes_0 = const()[name = string("op_221_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_23 = transpose(perm = var_172, x = var_171)[name = string("transpose_248")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_221 = expand_dims(axes = var_221_axes_0, x = x_23)[name = string("op_221")];
            tensor<int32, [5]> x_25_reps_0 = const()[name = string("x_25_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_25 = tile(reps = x_25_reps_0, x = var_221)[name = string("x_25")];
            tensor<int32, [4]> var_224 = const()[name = string("op_224"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_1 = reshape(shape = var_224, x = x_25)[name = string("value_1")];
            bool var_229_transpose_x_1 = const()[name = string("op_229_transpose_x_1"), val = bool(false)];
            bool var_229_transpose_y_1 = const()[name = string("op_229_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_229_cast_fp16 = matmul(transpose_x = var_229_transpose_x_1, transpose_y = var_229_transpose_y_1, x = query_1, y = key_1)[name = string("op_229_cast_fp16")];
            fp16 var_230_to_fp16 = const()[name = string("op_230_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_1_cast_fp16 = mul(x = var_229_cast_fp16, y = var_230_to_fp16)[name = string("attn_weights_1_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_3_cast_fp16 = add(x = attn_weights_1_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_234_cast_fp16 = softmax(axis = var_17, x = attn_weights_3_cast_fp16)[name = string("op_234_cast_fp16")];
            bool var_238_transpose_x_0 = const()[name = string("op_238_transpose_x_0"), val = bool(false)];
            bool var_238_transpose_y_0 = const()[name = string("op_238_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_238_cast_fp16 = matmul(transpose_x = var_238_transpose_x_0, transpose_y = var_238_transpose_y_0, x = var_234_cast_fp16, y = value_1)[name = string("op_238_cast_fp16")];
            tensor<int32, [4]> var_240 = const()[name = string("op_240"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_243 = const()[name = string("op_243"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_241 = transpose(perm = var_240, x = var_238_cast_fp16)[name = string("transpose_247")];
            tensor<fp16, [1, 1024, 2048]> attn_out_3 = reshape(shape = var_243, x = var_241)[name = string("attn_out_3")];
            tensor<int32, [3]> var_245 = const()[name = string("op_245"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_0 = const()[name = string("squeeze_0"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1091336512)))];
            string var_254_pad_type_0 = const()[name = string("op_254_pad_type_0"), val = string("valid")];
            int32 var_254_groups_0 = const()[name = string("op_254_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_254_strides_0 = const()[name = string("op_254_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_254_pad_0 = const()[name = string("op_254_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_254_dilations_0 = const()[name = string("op_254_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_246 = transpose(perm = var_245, x = attn_out_3)[name = string("transpose_246")];
            tensor<fp16, [1, 1024, 1024]> var_254 = conv(dilations = var_254_dilations_0, groups = var_254_groups_0, pad = var_254_pad_0, pad_type = var_254_pad_type_0, strides = var_254_strides_0, weight = squeeze_0, x = var_246)[name = string("op_254")];
            tensor<int32, [3]> var_255 = const()[name = string("op_255"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_5 = transpose(perm = var_255, x = var_254)[name = string("transpose_245")];
            tensor<fp16, [1, 1024, 1024]> x_27_cast_fp16 = add(x = var_82_cast_fp16, y = attn_out_5)[name = string("x_27_cast_fp16")];
            fp16 var_5_promoted_3_to_fp16 = const()[name = string("op_5_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_261_cast_fp16 = pow(x = x_27_cast_fp16, y = var_5_promoted_3_to_fp16)[name = string("op_261_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_261_cast_fp16)[name = string("var_7_cast_fp16")];
            fp16 var_264_to_fp16 = const()[name = string("op_264_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_265_cast_fp16 = add(x = var_7_cast_fp16, y = var_264_to_fp16)[name = string("op_265_cast_fp16")];
            fp32 var_266_epsilon_0 = const()[name = string("op_266_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_266_cast_fp16 = rsqrt(epsilon = var_266_epsilon_0, x = var_265_cast_fp16)[name = string("op_266_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_31_cast_fp16 = mul(x = x_27_cast_fp16, y = var_266_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_0_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_0_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1095530880)))];
            tensor<fp16, [1, 1024, 1024]> var_269_cast_fp16 = mul(x = x_31_cast_fp16, y = encoder_layers_0_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_269_cast_fp16")];
            tensor<int32, [3]> var_274 = const()[name = string("op_274"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_5_axes_0 = const()[name = string("input_5_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_275 = transpose(perm = var_274, x = var_269_cast_fp16)[name = string("transpose_244")];
            tensor<fp16, [1, 1024, 1, 1024]> input_5 = expand_dims(axes = input_5_axes_0, x = var_275)[name = string("input_5")];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_7 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = encoder_layers_0_mlp_gate_proj_weight, x = input_5)[name = string("input_7")];
            string up_1_pad_type_0 = const()[name = string("up_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_1_strides_0 = const()[name = string("up_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_1_pad_0 = const()[name = string("up_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_1_dilations_0 = const()[name = string("up_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_1_groups_0 = const()[name = string("up_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_1 = conv(dilations = up_1_dilations_0, groups = up_1_groups_0, pad = up_1_pad_0, pad_type = up_1_pad_type_0, strides = up_1_strides_0, weight = encoder_layers_0_mlp_up_proj_weight, x = input_5)[name = string("up_1")];
            tensor<fp16, [1, 3072, 1, 1024]> var_289 = silu(x = input_7)[name = string("op_289")];
            tensor<fp16, [1, 3072, 1, 1024]> input_9 = mul(x = var_289, y = up_1)[name = string("input_9")];
            string var_296_pad_type_0 = const()[name = string("op_296_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_296_strides_0 = const()[name = string("op_296_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_296_pad_0 = const()[name = string("op_296_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_296_dilations_0 = const()[name = string("op_296_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_296_groups_0 = const()[name = string("op_296_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_296 = conv(dilations = var_296_dilations_0, groups = var_296_groups_0, pad = var_296_pad_0, pad_type = var_296_pad_type_0, strides = var_296_strides_0, weight = encoder_layers_0_mlp_down_proj_weight, x = input_9)[name = string("op_296")];
            tensor<int32, [1]> var_297_axes_0 = const()[name = string("op_297_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_297 = squeeze(axes = var_297_axes_0, x = var_296)[name = string("op_297")];
            tensor<int32, [3]> var_298 = const()[name = string("op_298"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_1 = transpose(perm = var_298, x = var_297)[name = string("transpose_243")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_3_cast_fp16 = add(x = x_27_cast_fp16, y = mlp_out_1)[name = string("hidden_states_3_cast_fp16")];
            fp16 var_5_promoted_4_to_fp16 = const()[name = string("op_5_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_325_cast_fp16 = pow(x = hidden_states_3_cast_fp16, y = var_5_promoted_4_to_fp16)[name = string("op_325_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_325_cast_fp16)[name = string("var_9_cast_fp16")];
            fp16 var_328_to_fp16 = const()[name = string("op_328_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_329_cast_fp16 = add(x = var_9_cast_fp16, y = var_328_to_fp16)[name = string("op_329_cast_fp16")];
            fp32 var_330_epsilon_0 = const()[name = string("op_330_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_330_cast_fp16 = rsqrt(epsilon = var_330_epsilon_0, x = var_329_cast_fp16)[name = string("op_330_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_37_cast_fp16 = mul(x = hidden_states_3_cast_fp16, y = var_330_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_1_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_1_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1095532992)))];
            tensor<fp16, [1, 1024, 1024]> var_333_cast_fp16 = mul(x = x_37_cast_fp16, y = encoder_layers_1_input_layernorm_weight_promoted_to_fp16)[name = string("op_333_cast_fp16")];
            tensor<int32, [3]> var_338 = const()[name = string("op_338"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_11_axes_0 = const()[name = string("input_11_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_339 = transpose(perm = var_338, x = var_333_cast_fp16)[name = string("transpose_242")];
            tensor<fp16, [1, 1024, 1, 1024]> input_11 = expand_dims(axes = input_11_axes_0, x = var_339)[name = string("input_11")];
            string var_346_pad_type_0 = const()[name = string("op_346_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_346_strides_0 = const()[name = string("op_346_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_346_pad_0 = const()[name = string("op_346_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_346_dilations_0 = const()[name = string("op_346_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_346_groups_0 = const()[name = string("op_346_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_346 = conv(dilations = var_346_dilations_0, groups = var_346_groups_0, pad = var_346_pad_0, pad_type = var_346_pad_type_0, strides = var_346_strides_0, weight = encoder_layers_1_self_attn_q_proj_weight, x = input_11)[name = string("op_346")];
            tensor<int32, [4]> var_347 = const()[name = string("op_347"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_348 = reshape(shape = var_347, x = var_346)[name = string("op_348")];
            tensor<int32, [4]> var_349 = const()[name = string("op_349"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_356_pad_type_0 = const()[name = string("op_356_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_356_strides_0 = const()[name = string("op_356_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_356_pad_0 = const()[name = string("op_356_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_356_dilations_0 = const()[name = string("op_356_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_356_groups_0 = const()[name = string("op_356_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_356 = conv(dilations = var_356_dilations_0, groups = var_356_groups_0, pad = var_356_pad_0, pad_type = var_356_pad_type_0, strides = var_356_strides_0, weight = encoder_layers_1_self_attn_k_proj_weight, x = input_11)[name = string("op_356")];
            tensor<int32, [4]> var_357 = const()[name = string("op_357"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_358 = reshape(shape = var_357, x = var_356)[name = string("op_358")];
            tensor<int32, [4]> var_359 = const()[name = string("op_359"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_366_pad_type_0 = const()[name = string("op_366_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_366_strides_0 = const()[name = string("op_366_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_366_pad_0 = const()[name = string("op_366_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_366_dilations_0 = const()[name = string("op_366_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_366_groups_0 = const()[name = string("op_366_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_366 = conv(dilations = var_366_dilations_0, groups = var_366_groups_0, pad = var_366_pad_0, pad_type = var_366_pad_type_0, strides = var_366_strides_0, weight = encoder_layers_1_self_attn_v_proj_weight, x = input_11)[name = string("op_366")];
            tensor<int32, [4]> var_367 = const()[name = string("op_367"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_368 = reshape(shape = var_367, x = var_366)[name = string("op_368")];
            tensor<int32, [4]> var_369 = const()[name = string("op_369"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_5_to_fp16 = const()[name = string("op_5_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_7 = transpose(perm = var_349, x = var_348)[name = string("transpose_241")];
            tensor<fp16, [1, 16, 1024, 128]> var_375_cast_fp16 = pow(x = q_7, y = var_5_promoted_5_to_fp16)[name = string("op_375_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_375_cast_fp16)[name = string("var_11_cast_fp16")];
            fp16 var_378_to_fp16 = const()[name = string("op_378_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_379_cast_fp16 = add(x = var_11_cast_fp16, y = var_378_to_fp16)[name = string("op_379_cast_fp16")];
            fp32 var_380_epsilon_0 = const()[name = string("op_380_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_380_cast_fp16 = rsqrt(epsilon = var_380_epsilon_0, x = var_379_cast_fp16)[name = string("op_380_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_45_cast_fp16 = mul(x = q_7, y = var_380_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_9 = mul(x = x_45_cast_fp16, y = encoder_layers_1_self_attn_q_norm_weight)[name = string("q_9")];
            fp16 var_5_promoted_6_to_fp16 = const()[name = string("op_5_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_7 = transpose(perm = var_359, x = var_358)[name = string("transpose_240")];
            tensor<fp16, [1, 8, 1024, 128]> var_388_cast_fp16 = pow(x = k_7, y = var_5_promoted_6_to_fp16)[name = string("op_388_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_388_cast_fp16)[name = string("var_13_cast_fp16")];
            fp16 var_391_to_fp16 = const()[name = string("op_391_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_392_cast_fp16 = add(x = var_13_cast_fp16, y = var_391_to_fp16)[name = string("op_392_cast_fp16")];
            fp32 var_393_epsilon_0 = const()[name = string("op_393_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_393_cast_fp16 = rsqrt(epsilon = var_393_epsilon_0, x = var_392_cast_fp16)[name = string("op_393_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_51_cast_fp16 = mul(x = k_7, y = var_393_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_9 = mul(x = x_51_cast_fp16, y = encoder_layers_1_self_attn_k_norm_weight)[name = string("k_9")];
            tensor<fp16, [1, 16, 1024, 128]> var_397 = mul(x = q_9, y = cos)[name = string("op_397")];
            tensor<int32, [2]> var_398_split_sizes_0 = const()[name = string("op_398_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_398_axis_0 = const()[name = string("op_398_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_398_0, tensor<fp16, [1, 16, 1024, 64]> var_398_1 = split(axis = var_398_axis_0, split_sizes = var_398_split_sizes_0, x = q_9)[name = string("op_398")];
            fp16 const_6_promoted = const()[name = string("const_6_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_400 = mul(x = var_398_1, y = const_6_promoted)[name = string("op_400")];
            bool var_402_interleave_0 = const()[name = string("op_402_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_402 = concat(axis = var_17, interleave = var_402_interleave_0, values = (var_400, var_398_0))[name = string("op_402")];
            tensor<fp16, [1, 16, 1024, 128]> var_403 = mul(x = var_402, y = sin)[name = string("op_403")];
            tensor<fp16, [1, 16, 1024, 128]> query_3 = add(x = var_397, y = var_403)[name = string("query_3")];
            tensor<fp16, [1, 8, 1024, 128]> var_405 = mul(x = k_9, y = cos)[name = string("op_405")];
            tensor<int32, [2]> var_406_split_sizes_0 = const()[name = string("op_406_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_406_axis_0 = const()[name = string("op_406_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_406_0, tensor<fp16, [1, 8, 1024, 64]> var_406_1 = split(axis = var_406_axis_0, split_sizes = var_406_split_sizes_0, x = k_9)[name = string("op_406")];
            fp16 const_7_promoted = const()[name = string("const_7_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_408 = mul(x = var_406_1, y = const_7_promoted)[name = string("op_408")];
            bool var_410_interleave_0 = const()[name = string("op_410_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_410 = concat(axis = var_17, interleave = var_410_interleave_0, values = (var_408, var_406_0))[name = string("op_410")];
            tensor<fp16, [1, 8, 1024, 128]> var_411 = mul(x = var_410, y = sin)[name = string("op_411")];
            tensor<fp16, [1, 8, 1024, 128]> x_53 = add(x = var_405, y = var_411)[name = string("x_53")];
            tensor<int32, [1]> var_413_axes_0 = const()[name = string("op_413_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_413 = expand_dims(axes = var_413_axes_0, x = x_53)[name = string("op_413")];
            tensor<int32, [5]> x_55_reps_0 = const()[name = string("x_55_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_55 = tile(reps = x_55_reps_0, x = var_413)[name = string("x_55")];
            tensor<int32, [4]> var_416 = const()[name = string("op_416"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_3 = reshape(shape = var_416, x = x_55)[name = string("key_3")];
            tensor<int32, [1]> var_418_axes_0 = const()[name = string("op_418_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_57 = transpose(perm = var_369, x = var_368)[name = string("transpose_239")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_418 = expand_dims(axes = var_418_axes_0, x = x_57)[name = string("op_418")];
            tensor<int32, [5]> x_59_reps_0 = const()[name = string("x_59_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_59 = tile(reps = x_59_reps_0, x = var_418)[name = string("x_59")];
            tensor<int32, [4]> var_421 = const()[name = string("op_421"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_3 = reshape(shape = var_421, x = x_59)[name = string("value_3")];
            bool var_426_transpose_x_1 = const()[name = string("op_426_transpose_x_1"), val = bool(false)];
            bool var_426_transpose_y_1 = const()[name = string("op_426_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_426_cast_fp16 = matmul(transpose_x = var_426_transpose_x_1, transpose_y = var_426_transpose_y_1, x = query_3, y = key_3)[name = string("op_426_cast_fp16")];
            fp16 var_427_to_fp16 = const()[name = string("op_427_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_7_cast_fp16 = mul(x = var_426_cast_fp16, y = var_427_to_fp16)[name = string("attn_weights_7_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_9_cast_fp16 = add(x = attn_weights_7_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_9_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_431_cast_fp16 = softmax(axis = var_17, x = attn_weights_9_cast_fp16)[name = string("op_431_cast_fp16")];
            bool var_435_transpose_x_0 = const()[name = string("op_435_transpose_x_0"), val = bool(false)];
            bool var_435_transpose_y_0 = const()[name = string("op_435_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_435_cast_fp16 = matmul(transpose_x = var_435_transpose_x_0, transpose_y = var_435_transpose_y_0, x = var_431_cast_fp16, y = value_3)[name = string("op_435_cast_fp16")];
            tensor<int32, [4]> var_437 = const()[name = string("op_437"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_440 = const()[name = string("op_440"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_438 = transpose(perm = var_437, x = var_435_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [1, 1024, 2048]> attn_out_9 = reshape(shape = var_440, x = var_438)[name = string("attn_out_9")];
            tensor<int32, [3]> var_442 = const()[name = string("op_442"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_1 = const()[name = string("squeeze_1"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1095535104)))];
            string var_451_pad_type_0 = const()[name = string("op_451_pad_type_0"), val = string("valid")];
            int32 var_451_groups_0 = const()[name = string("op_451_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_451_strides_0 = const()[name = string("op_451_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_451_pad_0 = const()[name = string("op_451_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_451_dilations_0 = const()[name = string("op_451_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_443 = transpose(perm = var_442, x = attn_out_9)[name = string("transpose_237")];
            tensor<fp16, [1, 1024, 1024]> var_451 = conv(dilations = var_451_dilations_0, groups = var_451_groups_0, pad = var_451_pad_0, pad_type = var_451_pad_type_0, strides = var_451_strides_0, weight = squeeze_1, x = var_443)[name = string("op_451")];
            tensor<int32, [3]> var_452 = const()[name = string("op_452"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_11 = transpose(perm = var_452, x = var_451)[name = string("transpose_236")];
            tensor<fp16, [1, 1024, 1024]> x_61_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = attn_out_11)[name = string("x_61_cast_fp16")];
            fp16 var_5_promoted_7_to_fp16 = const()[name = string("op_5_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_458_cast_fp16 = pow(x = x_61_cast_fp16, y = var_5_promoted_7_to_fp16)[name = string("op_458_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_458_cast_fp16)[name = string("var_15_cast_fp16")];
            fp16 var_461_to_fp16 = const()[name = string("op_461_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_462_cast_fp16 = add(x = var_15_cast_fp16, y = var_461_to_fp16)[name = string("op_462_cast_fp16")];
            fp32 var_463_epsilon_0 = const()[name = string("op_463_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_463_cast_fp16 = rsqrt(epsilon = var_463_epsilon_0, x = var_462_cast_fp16)[name = string("op_463_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_65_cast_fp16 = mul(x = x_61_cast_fp16, y = var_463_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_1_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_1_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1099729472)))];
            tensor<fp16, [1, 1024, 1024]> var_466_cast_fp16 = mul(x = x_65_cast_fp16, y = encoder_layers_1_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_466_cast_fp16")];
            tensor<int32, [3]> var_471 = const()[name = string("op_471"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_472 = transpose(perm = var_471, x = var_466_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [1, 1024, 1, 1024]> input_15 = expand_dims(axes = input_15_axes_0, x = var_472)[name = string("input_15")];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_17 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = encoder_layers_1_mlp_gate_proj_weight, x = input_15)[name = string("input_17")];
            string up_3_pad_type_0 = const()[name = string("up_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_3_strides_0 = const()[name = string("up_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_3_pad_0 = const()[name = string("up_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_3_dilations_0 = const()[name = string("up_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_3_groups_0 = const()[name = string("up_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_3 = conv(dilations = up_3_dilations_0, groups = up_3_groups_0, pad = up_3_pad_0, pad_type = up_3_pad_type_0, strides = up_3_strides_0, weight = encoder_layers_1_mlp_up_proj_weight, x = input_15)[name = string("up_3")];
            tensor<fp16, [1, 3072, 1, 1024]> var_486 = silu(x = input_17)[name = string("op_486")];
            tensor<fp16, [1, 3072, 1, 1024]> input_19 = mul(x = var_486, y = up_3)[name = string("input_19")];
            string var_493_pad_type_0 = const()[name = string("op_493_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_493_strides_0 = const()[name = string("op_493_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_493_pad_0 = const()[name = string("op_493_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_493_dilations_0 = const()[name = string("op_493_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_493_groups_0 = const()[name = string("op_493_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_493 = conv(dilations = var_493_dilations_0, groups = var_493_groups_0, pad = var_493_pad_0, pad_type = var_493_pad_type_0, strides = var_493_strides_0, weight = encoder_layers_1_mlp_down_proj_weight, x = input_19)[name = string("op_493")];
            tensor<int32, [1]> var_494_axes_0 = const()[name = string("op_494_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_494 = squeeze(axes = var_494_axes_0, x = var_493)[name = string("op_494")];
            tensor<int32, [3]> var_495 = const()[name = string("op_495"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_3 = transpose(perm = var_495, x = var_494)[name = string("transpose_234")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_5_cast_fp16 = add(x = x_61_cast_fp16, y = mlp_out_3)[name = string("hidden_states_5_cast_fp16")];
            fp16 var_5_promoted_8_to_fp16 = const()[name = string("op_5_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_522_cast_fp16 = pow(x = hidden_states_5_cast_fp16, y = var_5_promoted_8_to_fp16)[name = string("op_522_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_522_cast_fp16)[name = string("var_17_cast_fp16")];
            fp16 var_525_to_fp16 = const()[name = string("op_525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_526_cast_fp16 = add(x = var_17_cast_fp16, y = var_525_to_fp16)[name = string("op_526_cast_fp16")];
            fp32 var_527_epsilon_0 = const()[name = string("op_527_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_527_cast_fp16 = rsqrt(epsilon = var_527_epsilon_0, x = var_526_cast_fp16)[name = string("op_527_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_71_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = var_527_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_2_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_2_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1099731584)))];
            tensor<fp16, [1, 1024, 1024]> var_530_cast_fp16 = mul(x = x_71_cast_fp16, y = encoder_layers_2_input_layernorm_weight_promoted_to_fp16)[name = string("op_530_cast_fp16")];
            tensor<int32, [3]> var_535 = const()[name = string("op_535"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_21_axes_0 = const()[name = string("input_21_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_536 = transpose(perm = var_535, x = var_530_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [1, 1024, 1, 1024]> input_21 = expand_dims(axes = input_21_axes_0, x = var_536)[name = string("input_21")];
            string var_543_pad_type_0 = const()[name = string("op_543_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_543_strides_0 = const()[name = string("op_543_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_543_pad_0 = const()[name = string("op_543_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_543_dilations_0 = const()[name = string("op_543_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_543_groups_0 = const()[name = string("op_543_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_543 = conv(dilations = var_543_dilations_0, groups = var_543_groups_0, pad = var_543_pad_0, pad_type = var_543_pad_type_0, strides = var_543_strides_0, weight = encoder_layers_2_self_attn_q_proj_weight, x = input_21)[name = string("op_543")];
            tensor<int32, [4]> var_544 = const()[name = string("op_544"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_545 = reshape(shape = var_544, x = var_543)[name = string("op_545")];
            tensor<int32, [4]> var_546 = const()[name = string("op_546"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_553_pad_type_0 = const()[name = string("op_553_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_553_strides_0 = const()[name = string("op_553_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_553_pad_0 = const()[name = string("op_553_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_553_dilations_0 = const()[name = string("op_553_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_553_groups_0 = const()[name = string("op_553_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_553 = conv(dilations = var_553_dilations_0, groups = var_553_groups_0, pad = var_553_pad_0, pad_type = var_553_pad_type_0, strides = var_553_strides_0, weight = encoder_layers_2_self_attn_k_proj_weight, x = input_21)[name = string("op_553")];
            tensor<int32, [4]> var_554 = const()[name = string("op_554"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_555 = reshape(shape = var_554, x = var_553)[name = string("op_555")];
            tensor<int32, [4]> var_556 = const()[name = string("op_556"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_563_pad_type_0 = const()[name = string("op_563_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_563_strides_0 = const()[name = string("op_563_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_563_pad_0 = const()[name = string("op_563_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_563_dilations_0 = const()[name = string("op_563_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_563_groups_0 = const()[name = string("op_563_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_563 = conv(dilations = var_563_dilations_0, groups = var_563_groups_0, pad = var_563_pad_0, pad_type = var_563_pad_type_0, strides = var_563_strides_0, weight = encoder_layers_2_self_attn_v_proj_weight, x = input_21)[name = string("op_563")];
            tensor<int32, [4]> var_564 = const()[name = string("op_564"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_565 = reshape(shape = var_564, x = var_563)[name = string("op_565")];
            tensor<int32, [4]> var_566 = const()[name = string("op_566"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_9_to_fp16 = const()[name = string("op_5_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_13 = transpose(perm = var_546, x = var_545)[name = string("transpose_232")];
            tensor<fp16, [1, 16, 1024, 128]> var_572_cast_fp16 = pow(x = q_13, y = var_5_promoted_9_to_fp16)[name = string("op_572_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_572_cast_fp16)[name = string("var_19_cast_fp16")];
            fp16 var_575_to_fp16 = const()[name = string("op_575_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_576_cast_fp16 = add(x = var_19_cast_fp16, y = var_575_to_fp16)[name = string("op_576_cast_fp16")];
            fp32 var_577_epsilon_0 = const()[name = string("op_577_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_577_cast_fp16 = rsqrt(epsilon = var_577_epsilon_0, x = var_576_cast_fp16)[name = string("op_577_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_79_cast_fp16 = mul(x = q_13, y = var_577_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_15 = mul(x = x_79_cast_fp16, y = encoder_layers_2_self_attn_q_norm_weight)[name = string("q_15")];
            fp16 var_5_promoted_10_to_fp16 = const()[name = string("op_5_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_13 = transpose(perm = var_556, x = var_555)[name = string("transpose_231")];
            tensor<fp16, [1, 8, 1024, 128]> var_585_cast_fp16 = pow(x = k_13, y = var_5_promoted_10_to_fp16)[name = string("op_585_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_585_cast_fp16)[name = string("var_21_cast_fp16")];
            fp16 var_588_to_fp16 = const()[name = string("op_588_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_589_cast_fp16 = add(x = var_21_cast_fp16, y = var_588_to_fp16)[name = string("op_589_cast_fp16")];
            fp32 var_590_epsilon_0 = const()[name = string("op_590_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_590_cast_fp16 = rsqrt(epsilon = var_590_epsilon_0, x = var_589_cast_fp16)[name = string("op_590_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_85_cast_fp16 = mul(x = k_13, y = var_590_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_15 = mul(x = x_85_cast_fp16, y = encoder_layers_2_self_attn_k_norm_weight)[name = string("k_15")];
            tensor<fp16, [1, 16, 1024, 128]> var_594 = mul(x = q_15, y = cos)[name = string("op_594")];
            tensor<int32, [2]> var_595_split_sizes_0 = const()[name = string("op_595_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_595_axis_0 = const()[name = string("op_595_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_595_0, tensor<fp16, [1, 16, 1024, 64]> var_595_1 = split(axis = var_595_axis_0, split_sizes = var_595_split_sizes_0, x = q_15)[name = string("op_595")];
            fp16 const_9_promoted = const()[name = string("const_9_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_597 = mul(x = var_595_1, y = const_9_promoted)[name = string("op_597")];
            bool var_599_interleave_0 = const()[name = string("op_599_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_599 = concat(axis = var_17, interleave = var_599_interleave_0, values = (var_597, var_595_0))[name = string("op_599")];
            tensor<fp16, [1, 16, 1024, 128]> var_600 = mul(x = var_599, y = sin)[name = string("op_600")];
            tensor<fp16, [1, 16, 1024, 128]> query_5 = add(x = var_594, y = var_600)[name = string("query_5")];
            tensor<fp16, [1, 8, 1024, 128]> var_602 = mul(x = k_15, y = cos)[name = string("op_602")];
            tensor<int32, [2]> var_603_split_sizes_0 = const()[name = string("op_603_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_603_axis_0 = const()[name = string("op_603_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_603_0, tensor<fp16, [1, 8, 1024, 64]> var_603_1 = split(axis = var_603_axis_0, split_sizes = var_603_split_sizes_0, x = k_15)[name = string("op_603")];
            fp16 const_10_promoted = const()[name = string("const_10_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_605 = mul(x = var_603_1, y = const_10_promoted)[name = string("op_605")];
            bool var_607_interleave_0 = const()[name = string("op_607_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_607 = concat(axis = var_17, interleave = var_607_interleave_0, values = (var_605, var_603_0))[name = string("op_607")];
            tensor<fp16, [1, 8, 1024, 128]> var_608 = mul(x = var_607, y = sin)[name = string("op_608")];
            tensor<fp16, [1, 8, 1024, 128]> x_87 = add(x = var_602, y = var_608)[name = string("x_87")];
            tensor<int32, [1]> var_610_axes_0 = const()[name = string("op_610_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_610 = expand_dims(axes = var_610_axes_0, x = x_87)[name = string("op_610")];
            tensor<int32, [5]> x_89_reps_0 = const()[name = string("x_89_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_89 = tile(reps = x_89_reps_0, x = var_610)[name = string("x_89")];
            tensor<int32, [4]> var_613 = const()[name = string("op_613"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_5 = reshape(shape = var_613, x = x_89)[name = string("key_5")];
            tensor<int32, [1]> var_615_axes_0 = const()[name = string("op_615_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_91 = transpose(perm = var_566, x = var_565)[name = string("transpose_230")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_615 = expand_dims(axes = var_615_axes_0, x = x_91)[name = string("op_615")];
            tensor<int32, [5]> x_93_reps_0 = const()[name = string("x_93_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_93 = tile(reps = x_93_reps_0, x = var_615)[name = string("x_93")];
            tensor<int32, [4]> var_618 = const()[name = string("op_618"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_5 = reshape(shape = var_618, x = x_93)[name = string("value_5")];
            bool var_623_transpose_x_1 = const()[name = string("op_623_transpose_x_1"), val = bool(false)];
            bool var_623_transpose_y_1 = const()[name = string("op_623_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_623_cast_fp16 = matmul(transpose_x = var_623_transpose_x_1, transpose_y = var_623_transpose_y_1, x = query_5, y = key_5)[name = string("op_623_cast_fp16")];
            fp16 var_624_to_fp16 = const()[name = string("op_624_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_13_cast_fp16 = mul(x = var_623_cast_fp16, y = var_624_to_fp16)[name = string("attn_weights_13_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_15_cast_fp16 = add(x = attn_weights_13_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_628_cast_fp16 = softmax(axis = var_17, x = attn_weights_15_cast_fp16)[name = string("op_628_cast_fp16")];
            bool var_632_transpose_x_0 = const()[name = string("op_632_transpose_x_0"), val = bool(false)];
            bool var_632_transpose_y_0 = const()[name = string("op_632_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_632_cast_fp16 = matmul(transpose_x = var_632_transpose_x_0, transpose_y = var_632_transpose_y_0, x = var_628_cast_fp16, y = value_5)[name = string("op_632_cast_fp16")];
            tensor<int32, [4]> var_634 = const()[name = string("op_634"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_637 = const()[name = string("op_637"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_635 = transpose(perm = var_634, x = var_632_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 1024, 2048]> attn_out_15 = reshape(shape = var_637, x = var_635)[name = string("attn_out_15")];
            tensor<int32, [3]> var_639 = const()[name = string("op_639"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_2 = const()[name = string("squeeze_2"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1099733696)))];
            string var_648_pad_type_0 = const()[name = string("op_648_pad_type_0"), val = string("valid")];
            int32 var_648_groups_0 = const()[name = string("op_648_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_648_strides_0 = const()[name = string("op_648_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_648_pad_0 = const()[name = string("op_648_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_648_dilations_0 = const()[name = string("op_648_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_640 = transpose(perm = var_639, x = attn_out_15)[name = string("transpose_228")];
            tensor<fp16, [1, 1024, 1024]> var_648 = conv(dilations = var_648_dilations_0, groups = var_648_groups_0, pad = var_648_pad_0, pad_type = var_648_pad_type_0, strides = var_648_strides_0, weight = squeeze_2, x = var_640)[name = string("op_648")];
            tensor<int32, [3]> var_649 = const()[name = string("op_649"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_17 = transpose(perm = var_649, x = var_648)[name = string("transpose_227")];
            tensor<fp16, [1, 1024, 1024]> x_95_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = attn_out_17)[name = string("x_95_cast_fp16")];
            fp16 var_5_promoted_11_to_fp16 = const()[name = string("op_5_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_655_cast_fp16 = pow(x = x_95_cast_fp16, y = var_5_promoted_11_to_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_655_cast_fp16)[name = string("var_23_cast_fp16")];
            fp16 var_658_to_fp16 = const()[name = string("op_658_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_659_cast_fp16 = add(x = var_23_cast_fp16, y = var_658_to_fp16)[name = string("op_659_cast_fp16")];
            fp32 var_660_epsilon_0 = const()[name = string("op_660_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_660_cast_fp16 = rsqrt(epsilon = var_660_epsilon_0, x = var_659_cast_fp16)[name = string("op_660_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_99_cast_fp16 = mul(x = x_95_cast_fp16, y = var_660_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_2_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_2_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1103928064)))];
            tensor<fp16, [1, 1024, 1024]> var_663_cast_fp16 = mul(x = x_99_cast_fp16, y = encoder_layers_2_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_663_cast_fp16")];
            tensor<int32, [3]> var_668 = const()[name = string("op_668"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_25_axes_0 = const()[name = string("input_25_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_669 = transpose(perm = var_668, x = var_663_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [1, 1024, 1, 1024]> input_25 = expand_dims(axes = input_25_axes_0, x = var_669)[name = string("input_25")];
            string input_27_pad_type_0 = const()[name = string("input_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_27_strides_0 = const()[name = string("input_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_27_pad_0 = const()[name = string("input_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_27_dilations_0 = const()[name = string("input_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_27_groups_0 = const()[name = string("input_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_27 = conv(dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = encoder_layers_2_mlp_gate_proj_weight, x = input_25)[name = string("input_27")];
            string up_5_pad_type_0 = const()[name = string("up_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_5_strides_0 = const()[name = string("up_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_5_pad_0 = const()[name = string("up_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_5_dilations_0 = const()[name = string("up_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_5_groups_0 = const()[name = string("up_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_5 = conv(dilations = up_5_dilations_0, groups = up_5_groups_0, pad = up_5_pad_0, pad_type = up_5_pad_type_0, strides = up_5_strides_0, weight = encoder_layers_2_mlp_up_proj_weight, x = input_25)[name = string("up_5")];
            tensor<fp16, [1, 3072, 1, 1024]> var_683 = silu(x = input_27)[name = string("op_683")];
            tensor<fp16, [1, 3072, 1, 1024]> input_29 = mul(x = var_683, y = up_5)[name = string("input_29")];
            string var_690_pad_type_0 = const()[name = string("op_690_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_690_strides_0 = const()[name = string("op_690_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_690_pad_0 = const()[name = string("op_690_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_690_dilations_0 = const()[name = string("op_690_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_690_groups_0 = const()[name = string("op_690_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_690 = conv(dilations = var_690_dilations_0, groups = var_690_groups_0, pad = var_690_pad_0, pad_type = var_690_pad_type_0, strides = var_690_strides_0, weight = encoder_layers_2_mlp_down_proj_weight, x = input_29)[name = string("op_690")];
            tensor<int32, [1]> var_691_axes_0 = const()[name = string("op_691_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_691 = squeeze(axes = var_691_axes_0, x = var_690)[name = string("op_691")];
            tensor<int32, [3]> var_692 = const()[name = string("op_692"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_5 = transpose(perm = var_692, x = var_691)[name = string("transpose_225")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_7_cast_fp16 = add(x = x_95_cast_fp16, y = mlp_out_5)[name = string("hidden_states_7_cast_fp16")];
            fp16 var_5_promoted_12_to_fp16 = const()[name = string("op_5_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_719_cast_fp16 = pow(x = hidden_states_7_cast_fp16, y = var_5_promoted_12_to_fp16)[name = string("op_719_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_719_cast_fp16)[name = string("var_25_cast_fp16")];
            fp16 var_722_to_fp16 = const()[name = string("op_722_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_723_cast_fp16 = add(x = var_25_cast_fp16, y = var_722_to_fp16)[name = string("op_723_cast_fp16")];
            fp32 var_724_epsilon_0 = const()[name = string("op_724_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_724_cast_fp16 = rsqrt(epsilon = var_724_epsilon_0, x = var_723_cast_fp16)[name = string("op_724_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_105_cast_fp16 = mul(x = hidden_states_7_cast_fp16, y = var_724_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_3_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_3_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1103930176)))];
            tensor<fp16, [1, 1024, 1024]> var_727_cast_fp16 = mul(x = x_105_cast_fp16, y = encoder_layers_3_input_layernorm_weight_promoted_to_fp16)[name = string("op_727_cast_fp16")];
            tensor<int32, [3]> var_732 = const()[name = string("op_732"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_31_axes_0 = const()[name = string("input_31_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_733 = transpose(perm = var_732, x = var_727_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [1, 1024, 1, 1024]> input_31 = expand_dims(axes = input_31_axes_0, x = var_733)[name = string("input_31")];
            string var_740_pad_type_0 = const()[name = string("op_740_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_740_strides_0 = const()[name = string("op_740_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_740_pad_0 = const()[name = string("op_740_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_740_dilations_0 = const()[name = string("op_740_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_740_groups_0 = const()[name = string("op_740_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_740 = conv(dilations = var_740_dilations_0, groups = var_740_groups_0, pad = var_740_pad_0, pad_type = var_740_pad_type_0, strides = var_740_strides_0, weight = encoder_layers_3_self_attn_q_proj_weight, x = input_31)[name = string("op_740")];
            tensor<int32, [4]> var_741 = const()[name = string("op_741"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_742 = reshape(shape = var_741, x = var_740)[name = string("op_742")];
            tensor<int32, [4]> var_743 = const()[name = string("op_743"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_750_pad_type_0 = const()[name = string("op_750_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_750_strides_0 = const()[name = string("op_750_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_750_pad_0 = const()[name = string("op_750_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_750_dilations_0 = const()[name = string("op_750_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_750_groups_0 = const()[name = string("op_750_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_750 = conv(dilations = var_750_dilations_0, groups = var_750_groups_0, pad = var_750_pad_0, pad_type = var_750_pad_type_0, strides = var_750_strides_0, weight = encoder_layers_3_self_attn_k_proj_weight, x = input_31)[name = string("op_750")];
            tensor<int32, [4]> var_751 = const()[name = string("op_751"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_752 = reshape(shape = var_751, x = var_750)[name = string("op_752")];
            tensor<int32, [4]> var_753 = const()[name = string("op_753"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_760_pad_type_0 = const()[name = string("op_760_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_760_strides_0 = const()[name = string("op_760_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_760_pad_0 = const()[name = string("op_760_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_760_dilations_0 = const()[name = string("op_760_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_760_groups_0 = const()[name = string("op_760_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_760 = conv(dilations = var_760_dilations_0, groups = var_760_groups_0, pad = var_760_pad_0, pad_type = var_760_pad_type_0, strides = var_760_strides_0, weight = encoder_layers_3_self_attn_v_proj_weight, x = input_31)[name = string("op_760")];
            tensor<int32, [4]> var_761 = const()[name = string("op_761"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_762 = reshape(shape = var_761, x = var_760)[name = string("op_762")];
            tensor<int32, [4]> var_763 = const()[name = string("op_763"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_13_to_fp16 = const()[name = string("op_5_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_19 = transpose(perm = var_743, x = var_742)[name = string("transpose_223")];
            tensor<fp16, [1, 16, 1024, 128]> var_769_cast_fp16 = pow(x = q_19, y = var_5_promoted_13_to_fp16)[name = string("op_769_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_769_cast_fp16)[name = string("var_27_cast_fp16")];
            fp16 var_772_to_fp16 = const()[name = string("op_772_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_773_cast_fp16 = add(x = var_27_cast_fp16, y = var_772_to_fp16)[name = string("op_773_cast_fp16")];
            fp32 var_774_epsilon_0 = const()[name = string("op_774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_774_cast_fp16 = rsqrt(epsilon = var_774_epsilon_0, x = var_773_cast_fp16)[name = string("op_774_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_113_cast_fp16 = mul(x = q_19, y = var_774_cast_fp16)[name = string("x_113_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_21 = mul(x = x_113_cast_fp16, y = encoder_layers_3_self_attn_q_norm_weight)[name = string("q_21")];
            fp16 var_5_promoted_14_to_fp16 = const()[name = string("op_5_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_19 = transpose(perm = var_753, x = var_752)[name = string("transpose_222")];
            tensor<fp16, [1, 8, 1024, 128]> var_782_cast_fp16 = pow(x = k_19, y = var_5_promoted_14_to_fp16)[name = string("op_782_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_782_cast_fp16)[name = string("var_29_cast_fp16")];
            fp16 var_785_to_fp16 = const()[name = string("op_785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_786_cast_fp16 = add(x = var_29_cast_fp16, y = var_785_to_fp16)[name = string("op_786_cast_fp16")];
            fp32 var_787_epsilon_0 = const()[name = string("op_787_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_787_cast_fp16 = rsqrt(epsilon = var_787_epsilon_0, x = var_786_cast_fp16)[name = string("op_787_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_119_cast_fp16 = mul(x = k_19, y = var_787_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_21 = mul(x = x_119_cast_fp16, y = encoder_layers_3_self_attn_k_norm_weight)[name = string("k_21")];
            tensor<fp16, [1, 16, 1024, 128]> var_791 = mul(x = q_21, y = cos)[name = string("op_791")];
            tensor<int32, [2]> var_792_split_sizes_0 = const()[name = string("op_792_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_792_axis_0 = const()[name = string("op_792_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_792_0, tensor<fp16, [1, 16, 1024, 64]> var_792_1 = split(axis = var_792_axis_0, split_sizes = var_792_split_sizes_0, x = q_21)[name = string("op_792")];
            fp16 const_12_promoted = const()[name = string("const_12_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_794 = mul(x = var_792_1, y = const_12_promoted)[name = string("op_794")];
            bool var_796_interleave_0 = const()[name = string("op_796_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_796 = concat(axis = var_17, interleave = var_796_interleave_0, values = (var_794, var_792_0))[name = string("op_796")];
            tensor<fp16, [1, 16, 1024, 128]> var_797 = mul(x = var_796, y = sin)[name = string("op_797")];
            tensor<fp16, [1, 16, 1024, 128]> query_7 = add(x = var_791, y = var_797)[name = string("query_7")];
            tensor<fp16, [1, 8, 1024, 128]> var_799 = mul(x = k_21, y = cos)[name = string("op_799")];
            tensor<int32, [2]> var_800_split_sizes_0 = const()[name = string("op_800_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_800_axis_0 = const()[name = string("op_800_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_800_0, tensor<fp16, [1, 8, 1024, 64]> var_800_1 = split(axis = var_800_axis_0, split_sizes = var_800_split_sizes_0, x = k_21)[name = string("op_800")];
            fp16 const_13_promoted = const()[name = string("const_13_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_802 = mul(x = var_800_1, y = const_13_promoted)[name = string("op_802")];
            bool var_804_interleave_0 = const()[name = string("op_804_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_804 = concat(axis = var_17, interleave = var_804_interleave_0, values = (var_802, var_800_0))[name = string("op_804")];
            tensor<fp16, [1, 8, 1024, 128]> var_805 = mul(x = var_804, y = sin)[name = string("op_805")];
            tensor<fp16, [1, 8, 1024, 128]> x_121 = add(x = var_799, y = var_805)[name = string("x_121")];
            tensor<int32, [1]> var_807_axes_0 = const()[name = string("op_807_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_807 = expand_dims(axes = var_807_axes_0, x = x_121)[name = string("op_807")];
            tensor<int32, [5]> x_123_reps_0 = const()[name = string("x_123_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_123 = tile(reps = x_123_reps_0, x = var_807)[name = string("x_123")];
            tensor<int32, [4]> var_810 = const()[name = string("op_810"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_7 = reshape(shape = var_810, x = x_123)[name = string("key_7")];
            tensor<int32, [1]> var_812_axes_0 = const()[name = string("op_812_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_125 = transpose(perm = var_763, x = var_762)[name = string("transpose_221")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_812 = expand_dims(axes = var_812_axes_0, x = x_125)[name = string("op_812")];
            tensor<int32, [5]> x_127_reps_0 = const()[name = string("x_127_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_127 = tile(reps = x_127_reps_0, x = var_812)[name = string("x_127")];
            tensor<int32, [4]> var_815 = const()[name = string("op_815"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_7 = reshape(shape = var_815, x = x_127)[name = string("value_7")];
            bool var_820_transpose_x_1 = const()[name = string("op_820_transpose_x_1"), val = bool(false)];
            bool var_820_transpose_y_1 = const()[name = string("op_820_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_820_cast_fp16 = matmul(transpose_x = var_820_transpose_x_1, transpose_y = var_820_transpose_y_1, x = query_7, y = key_7)[name = string("op_820_cast_fp16")];
            fp16 var_821_to_fp16 = const()[name = string("op_821_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_19_cast_fp16 = mul(x = var_820_cast_fp16, y = var_821_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_825_cast_fp16 = softmax(axis = var_17, x = attn_weights_21_cast_fp16)[name = string("op_825_cast_fp16")];
            bool var_829_transpose_x_0 = const()[name = string("op_829_transpose_x_0"), val = bool(false)];
            bool var_829_transpose_y_0 = const()[name = string("op_829_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_829_cast_fp16 = matmul(transpose_x = var_829_transpose_x_0, transpose_y = var_829_transpose_y_0, x = var_825_cast_fp16, y = value_7)[name = string("op_829_cast_fp16")];
            tensor<int32, [4]> var_831 = const()[name = string("op_831"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_834 = const()[name = string("op_834"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_832 = transpose(perm = var_831, x = var_829_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [1, 1024, 2048]> attn_out_21 = reshape(shape = var_834, x = var_832)[name = string("attn_out_21")];
            tensor<int32, [3]> var_836 = const()[name = string("op_836"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_3 = const()[name = string("squeeze_3"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1103932288)))];
            string var_845_pad_type_0 = const()[name = string("op_845_pad_type_0"), val = string("valid")];
            int32 var_845_groups_0 = const()[name = string("op_845_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_845_strides_0 = const()[name = string("op_845_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_845_pad_0 = const()[name = string("op_845_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_845_dilations_0 = const()[name = string("op_845_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_837 = transpose(perm = var_836, x = attn_out_21)[name = string("transpose_219")];
            tensor<fp16, [1, 1024, 1024]> var_845 = conv(dilations = var_845_dilations_0, groups = var_845_groups_0, pad = var_845_pad_0, pad_type = var_845_pad_type_0, strides = var_845_strides_0, weight = squeeze_3, x = var_837)[name = string("op_845")];
            tensor<int32, [3]> var_846 = const()[name = string("op_846"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_23 = transpose(perm = var_846, x = var_845)[name = string("transpose_218")];
            tensor<fp16, [1, 1024, 1024]> x_129_cast_fp16 = add(x = hidden_states_7_cast_fp16, y = attn_out_23)[name = string("x_129_cast_fp16")];
            fp16 var_5_promoted_15_to_fp16 = const()[name = string("op_5_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_852_cast_fp16 = pow(x = x_129_cast_fp16, y = var_5_promoted_15_to_fp16)[name = string("op_852_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_852_cast_fp16)[name = string("var_31_cast_fp16")];
            fp16 var_855_to_fp16 = const()[name = string("op_855_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_856_cast_fp16 = add(x = var_31_cast_fp16, y = var_855_to_fp16)[name = string("op_856_cast_fp16")];
            fp32 var_857_epsilon_0 = const()[name = string("op_857_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_857_cast_fp16 = rsqrt(epsilon = var_857_epsilon_0, x = var_856_cast_fp16)[name = string("op_857_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_133_cast_fp16 = mul(x = x_129_cast_fp16, y = var_857_cast_fp16)[name = string("x_133_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_3_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_3_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1108126656)))];
            tensor<fp16, [1, 1024, 1024]> var_860_cast_fp16 = mul(x = x_133_cast_fp16, y = encoder_layers_3_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [3]> var_865 = const()[name = string("op_865"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_866 = transpose(perm = var_865, x = var_860_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [1, 1024, 1, 1024]> input_35 = expand_dims(axes = input_35_axes_0, x = var_866)[name = string("input_35")];
            string input_37_pad_type_0 = const()[name = string("input_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_37_strides_0 = const()[name = string("input_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_37_pad_0 = const()[name = string("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_37_dilations_0 = const()[name = string("input_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_37_groups_0 = const()[name = string("input_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_37 = conv(dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = encoder_layers_3_mlp_gate_proj_weight, x = input_35)[name = string("input_37")];
            string up_7_pad_type_0 = const()[name = string("up_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_7_strides_0 = const()[name = string("up_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_7_pad_0 = const()[name = string("up_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_7_dilations_0 = const()[name = string("up_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_7_groups_0 = const()[name = string("up_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_7 = conv(dilations = up_7_dilations_0, groups = up_7_groups_0, pad = up_7_pad_0, pad_type = up_7_pad_type_0, strides = up_7_strides_0, weight = encoder_layers_3_mlp_up_proj_weight, x = input_35)[name = string("up_7")];
            tensor<fp16, [1, 3072, 1, 1024]> var_880 = silu(x = input_37)[name = string("op_880")];
            tensor<fp16, [1, 3072, 1, 1024]> input_39 = mul(x = var_880, y = up_7)[name = string("input_39")];
            string var_887_pad_type_0 = const()[name = string("op_887_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_887_strides_0 = const()[name = string("op_887_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_887_pad_0 = const()[name = string("op_887_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_887_dilations_0 = const()[name = string("op_887_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_887_groups_0 = const()[name = string("op_887_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_887 = conv(dilations = var_887_dilations_0, groups = var_887_groups_0, pad = var_887_pad_0, pad_type = var_887_pad_type_0, strides = var_887_strides_0, weight = encoder_layers_3_mlp_down_proj_weight, x = input_39)[name = string("op_887")];
            tensor<int32, [1]> var_888_axes_0 = const()[name = string("op_888_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_888 = squeeze(axes = var_888_axes_0, x = var_887)[name = string("op_888")];
            tensor<int32, [3]> var_889 = const()[name = string("op_889"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_7 = transpose(perm = var_889, x = var_888)[name = string("transpose_216")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_9_cast_fp16 = add(x = x_129_cast_fp16, y = mlp_out_7)[name = string("hidden_states_9_cast_fp16")];
            fp16 var_5_promoted_16_to_fp16 = const()[name = string("op_5_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_916_cast_fp16 = pow(x = hidden_states_9_cast_fp16, y = var_5_promoted_16_to_fp16)[name = string("op_916_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_916_cast_fp16)[name = string("var_33_cast_fp16")];
            fp16 var_919_to_fp16 = const()[name = string("op_919_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_920_cast_fp16 = add(x = var_33_cast_fp16, y = var_919_to_fp16)[name = string("op_920_cast_fp16")];
            fp32 var_921_epsilon_0 = const()[name = string("op_921_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_921_cast_fp16 = rsqrt(epsilon = var_921_epsilon_0, x = var_920_cast_fp16)[name = string("op_921_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_139_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = var_921_cast_fp16)[name = string("x_139_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_4_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_4_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1108128768)))];
            tensor<fp16, [1, 1024, 1024]> var_924_cast_fp16 = mul(x = x_139_cast_fp16, y = encoder_layers_4_input_layernorm_weight_promoted_to_fp16)[name = string("op_924_cast_fp16")];
            tensor<int32, [3]> var_929 = const()[name = string("op_929"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_41_axes_0 = const()[name = string("input_41_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_930 = transpose(perm = var_929, x = var_924_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [1, 1024, 1, 1024]> input_41 = expand_dims(axes = input_41_axes_0, x = var_930)[name = string("input_41")];
            string var_937_pad_type_0 = const()[name = string("op_937_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_937_strides_0 = const()[name = string("op_937_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_937_pad_0 = const()[name = string("op_937_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_937_dilations_0 = const()[name = string("op_937_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_937_groups_0 = const()[name = string("op_937_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_937 = conv(dilations = var_937_dilations_0, groups = var_937_groups_0, pad = var_937_pad_0, pad_type = var_937_pad_type_0, strides = var_937_strides_0, weight = encoder_layers_4_self_attn_q_proj_weight, x = input_41)[name = string("op_937")];
            tensor<int32, [4]> var_938 = const()[name = string("op_938"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_939 = reshape(shape = var_938, x = var_937)[name = string("op_939")];
            tensor<int32, [4]> var_940 = const()[name = string("op_940"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_947_pad_type_0 = const()[name = string("op_947_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_947_strides_0 = const()[name = string("op_947_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_947_pad_0 = const()[name = string("op_947_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_947_dilations_0 = const()[name = string("op_947_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_947_groups_0 = const()[name = string("op_947_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_947 = conv(dilations = var_947_dilations_0, groups = var_947_groups_0, pad = var_947_pad_0, pad_type = var_947_pad_type_0, strides = var_947_strides_0, weight = encoder_layers_4_self_attn_k_proj_weight, x = input_41)[name = string("op_947")];
            tensor<int32, [4]> var_948 = const()[name = string("op_948"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_949 = reshape(shape = var_948, x = var_947)[name = string("op_949")];
            tensor<int32, [4]> var_950 = const()[name = string("op_950"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_957_pad_type_0 = const()[name = string("op_957_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_957_strides_0 = const()[name = string("op_957_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_957_pad_0 = const()[name = string("op_957_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_957_dilations_0 = const()[name = string("op_957_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_957_groups_0 = const()[name = string("op_957_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_957 = conv(dilations = var_957_dilations_0, groups = var_957_groups_0, pad = var_957_pad_0, pad_type = var_957_pad_type_0, strides = var_957_strides_0, weight = encoder_layers_4_self_attn_v_proj_weight, x = input_41)[name = string("op_957")];
            tensor<int32, [4]> var_958 = const()[name = string("op_958"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_959 = reshape(shape = var_958, x = var_957)[name = string("op_959")];
            tensor<int32, [4]> var_960 = const()[name = string("op_960"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_17_to_fp16 = const()[name = string("op_5_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_25 = transpose(perm = var_940, x = var_939)[name = string("transpose_214")];
            tensor<fp16, [1, 16, 1024, 128]> var_966_cast_fp16 = pow(x = q_25, y = var_5_promoted_17_to_fp16)[name = string("op_966_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_966_cast_fp16)[name = string("var_35_cast_fp16")];
            fp16 var_969_to_fp16 = const()[name = string("op_969_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_970_cast_fp16 = add(x = var_35_cast_fp16, y = var_969_to_fp16)[name = string("op_970_cast_fp16")];
            fp32 var_971_epsilon_0 = const()[name = string("op_971_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_971_cast_fp16 = rsqrt(epsilon = var_971_epsilon_0, x = var_970_cast_fp16)[name = string("op_971_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_147_cast_fp16 = mul(x = q_25, y = var_971_cast_fp16)[name = string("x_147_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_27 = mul(x = x_147_cast_fp16, y = encoder_layers_4_self_attn_q_norm_weight)[name = string("q_27")];
            fp16 var_5_promoted_18_to_fp16 = const()[name = string("op_5_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_25 = transpose(perm = var_950, x = var_949)[name = string("transpose_213")];
            tensor<fp16, [1, 8, 1024, 128]> var_979_cast_fp16 = pow(x = k_25, y = var_5_promoted_18_to_fp16)[name = string("op_979_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_979_cast_fp16)[name = string("var_37_cast_fp16")];
            fp16 var_982_to_fp16 = const()[name = string("op_982_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_983_cast_fp16 = add(x = var_37_cast_fp16, y = var_982_to_fp16)[name = string("op_983_cast_fp16")];
            fp32 var_984_epsilon_0 = const()[name = string("op_984_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_984_cast_fp16 = rsqrt(epsilon = var_984_epsilon_0, x = var_983_cast_fp16)[name = string("op_984_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_153_cast_fp16 = mul(x = k_25, y = var_984_cast_fp16)[name = string("x_153_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_27 = mul(x = x_153_cast_fp16, y = encoder_layers_4_self_attn_k_norm_weight)[name = string("k_27")];
            tensor<fp16, [1, 16, 1024, 128]> var_988 = mul(x = q_27, y = cos)[name = string("op_988")];
            tensor<int32, [2]> var_989_split_sizes_0 = const()[name = string("op_989_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_989_axis_0 = const()[name = string("op_989_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_989_0, tensor<fp16, [1, 16, 1024, 64]> var_989_1 = split(axis = var_989_axis_0, split_sizes = var_989_split_sizes_0, x = q_27)[name = string("op_989")];
            fp16 const_15_promoted = const()[name = string("const_15_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_991 = mul(x = var_989_1, y = const_15_promoted)[name = string("op_991")];
            bool var_993_interleave_0 = const()[name = string("op_993_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_993 = concat(axis = var_17, interleave = var_993_interleave_0, values = (var_991, var_989_0))[name = string("op_993")];
            tensor<fp16, [1, 16, 1024, 128]> var_994 = mul(x = var_993, y = sin)[name = string("op_994")];
            tensor<fp16, [1, 16, 1024, 128]> query_9 = add(x = var_988, y = var_994)[name = string("query_9")];
            tensor<fp16, [1, 8, 1024, 128]> var_996 = mul(x = k_27, y = cos)[name = string("op_996")];
            tensor<int32, [2]> var_997_split_sizes_0 = const()[name = string("op_997_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_997_axis_0 = const()[name = string("op_997_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_997_0, tensor<fp16, [1, 8, 1024, 64]> var_997_1 = split(axis = var_997_axis_0, split_sizes = var_997_split_sizes_0, x = k_27)[name = string("op_997")];
            fp16 const_16_promoted = const()[name = string("const_16_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_999 = mul(x = var_997_1, y = const_16_promoted)[name = string("op_999")];
            bool var_1001_interleave_0 = const()[name = string("op_1001_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_1001 = concat(axis = var_17, interleave = var_1001_interleave_0, values = (var_999, var_997_0))[name = string("op_1001")];
            tensor<fp16, [1, 8, 1024, 128]> var_1002 = mul(x = var_1001, y = sin)[name = string("op_1002")];
            tensor<fp16, [1, 8, 1024, 128]> x_155 = add(x = var_996, y = var_1002)[name = string("x_155")];
            tensor<int32, [1]> var_1004_axes_0 = const()[name = string("op_1004_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1004 = expand_dims(axes = var_1004_axes_0, x = x_155)[name = string("op_1004")];
            tensor<int32, [5]> x_157_reps_0 = const()[name = string("x_157_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_157 = tile(reps = x_157_reps_0, x = var_1004)[name = string("x_157")];
            tensor<int32, [4]> var_1007 = const()[name = string("op_1007"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_9 = reshape(shape = var_1007, x = x_157)[name = string("key_9")];
            tensor<int32, [1]> var_1009_axes_0 = const()[name = string("op_1009_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_159 = transpose(perm = var_960, x = var_959)[name = string("transpose_212")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1009 = expand_dims(axes = var_1009_axes_0, x = x_159)[name = string("op_1009")];
            tensor<int32, [5]> x_161_reps_0 = const()[name = string("x_161_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_161 = tile(reps = x_161_reps_0, x = var_1009)[name = string("x_161")];
            tensor<int32, [4]> var_1012 = const()[name = string("op_1012"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_9 = reshape(shape = var_1012, x = x_161)[name = string("value_9")];
            bool var_1017_transpose_x_1 = const()[name = string("op_1017_transpose_x_1"), val = bool(false)];
            bool var_1017_transpose_y_1 = const()[name = string("op_1017_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_1017_cast_fp16 = matmul(transpose_x = var_1017_transpose_x_1, transpose_y = var_1017_transpose_y_1, x = query_9, y = key_9)[name = string("op_1017_cast_fp16")];
            fp16 var_1018_to_fp16 = const()[name = string("op_1018_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_25_cast_fp16 = mul(x = var_1017_cast_fp16, y = var_1018_to_fp16)[name = string("attn_weights_25_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_27_cast_fp16 = add(x = attn_weights_25_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1022_cast_fp16 = softmax(axis = var_17, x = attn_weights_27_cast_fp16)[name = string("op_1022_cast_fp16")];
            bool var_1026_transpose_x_0 = const()[name = string("op_1026_transpose_x_0"), val = bool(false)];
            bool var_1026_transpose_y_0 = const()[name = string("op_1026_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_1026_cast_fp16 = matmul(transpose_x = var_1026_transpose_x_0, transpose_y = var_1026_transpose_y_0, x = var_1022_cast_fp16, y = value_9)[name = string("op_1026_cast_fp16")];
            tensor<int32, [4]> var_1028 = const()[name = string("op_1028"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1031 = const()[name = string("op_1031"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_1029 = transpose(perm = var_1028, x = var_1026_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [1, 1024, 2048]> attn_out_27 = reshape(shape = var_1031, x = var_1029)[name = string("attn_out_27")];
            tensor<int32, [3]> var_1033 = const()[name = string("op_1033"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_4 = const()[name = string("squeeze_4"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1108130880)))];
            string var_1042_pad_type_0 = const()[name = string("op_1042_pad_type_0"), val = string("valid")];
            int32 var_1042_groups_0 = const()[name = string("op_1042_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_1042_strides_0 = const()[name = string("op_1042_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1042_pad_0 = const()[name = string("op_1042_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1042_dilations_0 = const()[name = string("op_1042_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_1034 = transpose(perm = var_1033, x = attn_out_27)[name = string("transpose_210")];
            tensor<fp16, [1, 1024, 1024]> var_1042 = conv(dilations = var_1042_dilations_0, groups = var_1042_groups_0, pad = var_1042_pad_0, pad_type = var_1042_pad_type_0, strides = var_1042_strides_0, weight = squeeze_4, x = var_1034)[name = string("op_1042")];
            tensor<int32, [3]> var_1043 = const()[name = string("op_1043"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_29 = transpose(perm = var_1043, x = var_1042)[name = string("transpose_209")];
            tensor<fp16, [1, 1024, 1024]> x_163_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = attn_out_29)[name = string("x_163_cast_fp16")];
            fp16 var_5_promoted_19_to_fp16 = const()[name = string("op_5_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_1049_cast_fp16 = pow(x = x_163_cast_fp16, y = var_5_promoted_19_to_fp16)[name = string("op_1049_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1049_cast_fp16)[name = string("var_39_cast_fp16")];
            fp16 var_1052_to_fp16 = const()[name = string("op_1052_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1053_cast_fp16 = add(x = var_39_cast_fp16, y = var_1052_to_fp16)[name = string("op_1053_cast_fp16")];
            fp32 var_1054_epsilon_0 = const()[name = string("op_1054_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1054_cast_fp16 = rsqrt(epsilon = var_1054_epsilon_0, x = var_1053_cast_fp16)[name = string("op_1054_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_167_cast_fp16 = mul(x = x_163_cast_fp16, y = var_1054_cast_fp16)[name = string("x_167_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_4_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_4_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1112325248)))];
            tensor<fp16, [1, 1024, 1024]> var_1057_cast_fp16 = mul(x = x_167_cast_fp16, y = encoder_layers_4_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_1057_cast_fp16")];
            tensor<int32, [3]> var_1062 = const()[name = string("op_1062"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_45_axes_0 = const()[name = string("input_45_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1063 = transpose(perm = var_1062, x = var_1057_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 1024, 1, 1024]> input_45 = expand_dims(axes = input_45_axes_0, x = var_1063)[name = string("input_45")];
            string input_47_pad_type_0 = const()[name = string("input_47_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_47_strides_0 = const()[name = string("input_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_47_pad_0 = const()[name = string("input_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_47_dilations_0 = const()[name = string("input_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_47_groups_0 = const()[name = string("input_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_47 = conv(dilations = input_47_dilations_0, groups = input_47_groups_0, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = input_47_strides_0, weight = encoder_layers_4_mlp_gate_proj_weight, x = input_45)[name = string("input_47")];
            string up_9_pad_type_0 = const()[name = string("up_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_9_strides_0 = const()[name = string("up_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_9_pad_0 = const()[name = string("up_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_9_dilations_0 = const()[name = string("up_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_9_groups_0 = const()[name = string("up_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_9 = conv(dilations = up_9_dilations_0, groups = up_9_groups_0, pad = up_9_pad_0, pad_type = up_9_pad_type_0, strides = up_9_strides_0, weight = encoder_layers_4_mlp_up_proj_weight, x = input_45)[name = string("up_9")];
            tensor<fp16, [1, 3072, 1, 1024]> var_1077 = silu(x = input_47)[name = string("op_1077")];
            tensor<fp16, [1, 3072, 1, 1024]> input_49 = mul(x = var_1077, y = up_9)[name = string("input_49")];
            string var_1084_pad_type_0 = const()[name = string("op_1084_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1084_strides_0 = const()[name = string("op_1084_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1084_pad_0 = const()[name = string("op_1084_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1084_dilations_0 = const()[name = string("op_1084_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1084_groups_0 = const()[name = string("op_1084_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1084 = conv(dilations = var_1084_dilations_0, groups = var_1084_groups_0, pad = var_1084_pad_0, pad_type = var_1084_pad_type_0, strides = var_1084_strides_0, weight = encoder_layers_4_mlp_down_proj_weight, x = input_49)[name = string("op_1084")];
            tensor<int32, [1]> var_1085_axes_0 = const()[name = string("op_1085_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1085 = squeeze(axes = var_1085_axes_0, x = var_1084)[name = string("op_1085")];
            tensor<int32, [3]> var_1086 = const()[name = string("op_1086"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_9 = transpose(perm = var_1086, x = var_1085)[name = string("transpose_207")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_11_cast_fp16 = add(x = x_163_cast_fp16, y = mlp_out_9)[name = string("hidden_states_11_cast_fp16")];
            fp16 var_5_promoted_20_to_fp16 = const()[name = string("op_5_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_1113_cast_fp16 = pow(x = hidden_states_11_cast_fp16, y = var_5_promoted_20_to_fp16)[name = string("op_1113_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1113_cast_fp16)[name = string("var_41_cast_fp16")];
            fp16 var_1116_to_fp16 = const()[name = string("op_1116_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1117_cast_fp16 = add(x = var_41_cast_fp16, y = var_1116_to_fp16)[name = string("op_1117_cast_fp16")];
            fp32 var_1118_epsilon_0 = const()[name = string("op_1118_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1118_cast_fp16 = rsqrt(epsilon = var_1118_epsilon_0, x = var_1117_cast_fp16)[name = string("op_1118_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_173_cast_fp16 = mul(x = hidden_states_11_cast_fp16, y = var_1118_cast_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_5_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_5_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1112327360)))];
            tensor<fp16, [1, 1024, 1024]> var_1121_cast_fp16 = mul(x = x_173_cast_fp16, y = encoder_layers_5_input_layernorm_weight_promoted_to_fp16)[name = string("op_1121_cast_fp16")];
            tensor<int32, [3]> var_1126 = const()[name = string("op_1126"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_51_axes_0 = const()[name = string("input_51_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1127 = transpose(perm = var_1126, x = var_1121_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [1, 1024, 1, 1024]> input_51 = expand_dims(axes = input_51_axes_0, x = var_1127)[name = string("input_51")];
            string var_1134_pad_type_0 = const()[name = string("op_1134_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1134_strides_0 = const()[name = string("op_1134_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1134_pad_0 = const()[name = string("op_1134_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1134_dilations_0 = const()[name = string("op_1134_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1134_groups_0 = const()[name = string("op_1134_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_1134 = conv(dilations = var_1134_dilations_0, groups = var_1134_groups_0, pad = var_1134_pad_0, pad_type = var_1134_pad_type_0, strides = var_1134_strides_0, weight = encoder_layers_5_self_attn_q_proj_weight, x = input_51)[name = string("op_1134")];
            tensor<int32, [4]> var_1135 = const()[name = string("op_1135"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_1136 = reshape(shape = var_1135, x = var_1134)[name = string("op_1136")];
            tensor<int32, [4]> var_1137 = const()[name = string("op_1137"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1144_pad_type_0 = const()[name = string("op_1144_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1144_strides_0 = const()[name = string("op_1144_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1144_pad_0 = const()[name = string("op_1144_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1144_dilations_0 = const()[name = string("op_1144_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1144_groups_0 = const()[name = string("op_1144_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1144 = conv(dilations = var_1144_dilations_0, groups = var_1144_groups_0, pad = var_1144_pad_0, pad_type = var_1144_pad_type_0, strides = var_1144_strides_0, weight = encoder_layers_5_self_attn_k_proj_weight, x = input_51)[name = string("op_1144")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_1146 = reshape(shape = var_1145, x = var_1144)[name = string("op_1146")];
            tensor<int32, [4]> var_1147 = const()[name = string("op_1147"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1154_pad_type_0 = const()[name = string("op_1154_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1154_strides_0 = const()[name = string("op_1154_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1154_pad_0 = const()[name = string("op_1154_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1154_dilations_0 = const()[name = string("op_1154_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1154_groups_0 = const()[name = string("op_1154_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1154 = conv(dilations = var_1154_dilations_0, groups = var_1154_groups_0, pad = var_1154_pad_0, pad_type = var_1154_pad_type_0, strides = var_1154_strides_0, weight = encoder_layers_5_self_attn_v_proj_weight, x = input_51)[name = string("op_1154")];
            tensor<int32, [4]> var_1155 = const()[name = string("op_1155"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_1156 = reshape(shape = var_1155, x = var_1154)[name = string("op_1156")];
            tensor<int32, [4]> var_1157 = const()[name = string("op_1157"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_21_to_fp16 = const()[name = string("op_5_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_31 = transpose(perm = var_1137, x = var_1136)[name = string("transpose_205")];
            tensor<fp16, [1, 16, 1024, 128]> var_1163_cast_fp16 = pow(x = q_31, y = var_5_promoted_21_to_fp16)[name = string("op_1163_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1163_cast_fp16)[name = string("var_43_cast_fp16")];
            fp16 var_1166_to_fp16 = const()[name = string("op_1166_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_1167_cast_fp16 = add(x = var_43_cast_fp16, y = var_1166_to_fp16)[name = string("op_1167_cast_fp16")];
            fp32 var_1168_epsilon_0 = const()[name = string("op_1168_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_1168_cast_fp16 = rsqrt(epsilon = var_1168_epsilon_0, x = var_1167_cast_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_181_cast_fp16 = mul(x = q_31, y = var_1168_cast_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_33 = mul(x = x_181_cast_fp16, y = encoder_layers_5_self_attn_q_norm_weight)[name = string("q_33")];
            fp16 var_5_promoted_22_to_fp16 = const()[name = string("op_5_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_31 = transpose(perm = var_1147, x = var_1146)[name = string("transpose_204")];
            tensor<fp16, [1, 8, 1024, 128]> var_1176_cast_fp16 = pow(x = k_31, y = var_5_promoted_22_to_fp16)[name = string("op_1176_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1176_cast_fp16)[name = string("var_45_cast_fp16")];
            fp16 var_1179_to_fp16 = const()[name = string("op_1179_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_1180_cast_fp16 = add(x = var_45_cast_fp16, y = var_1179_to_fp16)[name = string("op_1180_cast_fp16")];
            fp32 var_1181_epsilon_0 = const()[name = string("op_1181_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_1181_cast_fp16 = rsqrt(epsilon = var_1181_epsilon_0, x = var_1180_cast_fp16)[name = string("op_1181_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_187_cast_fp16 = mul(x = k_31, y = var_1181_cast_fp16)[name = string("x_187_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_33 = mul(x = x_187_cast_fp16, y = encoder_layers_5_self_attn_k_norm_weight)[name = string("k_33")];
            tensor<fp16, [1, 16, 1024, 128]> var_1185 = mul(x = q_33, y = cos)[name = string("op_1185")];
            tensor<int32, [2]> var_1186_split_sizes_0 = const()[name = string("op_1186_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1186_axis_0 = const()[name = string("op_1186_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_1186_0, tensor<fp16, [1, 16, 1024, 64]> var_1186_1 = split(axis = var_1186_axis_0, split_sizes = var_1186_split_sizes_0, x = q_33)[name = string("op_1186")];
            fp16 const_18_promoted = const()[name = string("const_18_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_1188 = mul(x = var_1186_1, y = const_18_promoted)[name = string("op_1188")];
            bool var_1190_interleave_0 = const()[name = string("op_1190_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_1190 = concat(axis = var_17, interleave = var_1190_interleave_0, values = (var_1188, var_1186_0))[name = string("op_1190")];
            tensor<fp16, [1, 16, 1024, 128]> var_1191 = mul(x = var_1190, y = sin)[name = string("op_1191")];
            tensor<fp16, [1, 16, 1024, 128]> query_11 = add(x = var_1185, y = var_1191)[name = string("query_11")];
            tensor<fp16, [1, 8, 1024, 128]> var_1193 = mul(x = k_33, y = cos)[name = string("op_1193")];
            tensor<int32, [2]> var_1194_split_sizes_0 = const()[name = string("op_1194_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1194_axis_0 = const()[name = string("op_1194_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_1194_0, tensor<fp16, [1, 8, 1024, 64]> var_1194_1 = split(axis = var_1194_axis_0, split_sizes = var_1194_split_sizes_0, x = k_33)[name = string("op_1194")];
            fp16 const_19_promoted = const()[name = string("const_19_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_1196 = mul(x = var_1194_1, y = const_19_promoted)[name = string("op_1196")];
            bool var_1198_interleave_0 = const()[name = string("op_1198_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_1198 = concat(axis = var_17, interleave = var_1198_interleave_0, values = (var_1196, var_1194_0))[name = string("op_1198")];
            tensor<fp16, [1, 8, 1024, 128]> var_1199 = mul(x = var_1198, y = sin)[name = string("op_1199")];
            tensor<fp16, [1, 8, 1024, 128]> x_189 = add(x = var_1193, y = var_1199)[name = string("x_189")];
            tensor<int32, [1]> var_1201_axes_0 = const()[name = string("op_1201_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1201 = expand_dims(axes = var_1201_axes_0, x = x_189)[name = string("op_1201")];
            tensor<int32, [5]> x_191_reps_0 = const()[name = string("x_191_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_191 = tile(reps = x_191_reps_0, x = var_1201)[name = string("x_191")];
            tensor<int32, [4]> var_1204 = const()[name = string("op_1204"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_11 = reshape(shape = var_1204, x = x_191)[name = string("key_11")];
            tensor<int32, [1]> var_1206_axes_0 = const()[name = string("op_1206_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_193 = transpose(perm = var_1157, x = var_1156)[name = string("transpose_203")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1206 = expand_dims(axes = var_1206_axes_0, x = x_193)[name = string("op_1206")];
            tensor<int32, [5]> x_195_reps_0 = const()[name = string("x_195_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_195 = tile(reps = x_195_reps_0, x = var_1206)[name = string("x_195")];
            tensor<int32, [4]> var_1209 = const()[name = string("op_1209"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_11 = reshape(shape = var_1209, x = x_195)[name = string("value_11")];
            bool var_1214_transpose_x_1 = const()[name = string("op_1214_transpose_x_1"), val = bool(false)];
            bool var_1214_transpose_y_1 = const()[name = string("op_1214_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_1214_cast_fp16 = matmul(transpose_x = var_1214_transpose_x_1, transpose_y = var_1214_transpose_y_1, x = query_11, y = key_11)[name = string("op_1214_cast_fp16")];
            fp16 var_1215_to_fp16 = const()[name = string("op_1215_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_31_cast_fp16 = mul(x = var_1214_cast_fp16, y = var_1215_to_fp16)[name = string("attn_weights_31_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_33_cast_fp16 = add(x = attn_weights_31_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_33_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1219_cast_fp16 = softmax(axis = var_17, x = attn_weights_33_cast_fp16)[name = string("op_1219_cast_fp16")];
            bool var_1223_transpose_x_0 = const()[name = string("op_1223_transpose_x_0"), val = bool(false)];
            bool var_1223_transpose_y_0 = const()[name = string("op_1223_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_1223_cast_fp16 = matmul(transpose_x = var_1223_transpose_x_0, transpose_y = var_1223_transpose_y_0, x = var_1219_cast_fp16, y = value_11)[name = string("op_1223_cast_fp16")];
            tensor<int32, [4]> var_1225 = const()[name = string("op_1225"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1228 = const()[name = string("op_1228"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_1226 = transpose(perm = var_1225, x = var_1223_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [1, 1024, 2048]> attn_out_33 = reshape(shape = var_1228, x = var_1226)[name = string("attn_out_33")];
            tensor<int32, [3]> var_1230 = const()[name = string("op_1230"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_5 = const()[name = string("squeeze_5"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1112329472)))];
            string var_1239_pad_type_0 = const()[name = string("op_1239_pad_type_0"), val = string("valid")];
            int32 var_1239_groups_0 = const()[name = string("op_1239_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_1239_strides_0 = const()[name = string("op_1239_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1239_pad_0 = const()[name = string("op_1239_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1239_dilations_0 = const()[name = string("op_1239_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_1231 = transpose(perm = var_1230, x = attn_out_33)[name = string("transpose_201")];
            tensor<fp16, [1, 1024, 1024]> var_1239 = conv(dilations = var_1239_dilations_0, groups = var_1239_groups_0, pad = var_1239_pad_0, pad_type = var_1239_pad_type_0, strides = var_1239_strides_0, weight = squeeze_5, x = var_1231)[name = string("op_1239")];
            tensor<int32, [3]> var_1240 = const()[name = string("op_1240"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_35 = transpose(perm = var_1240, x = var_1239)[name = string("transpose_200")];
            tensor<fp16, [1, 1024, 1024]> x_197_cast_fp16 = add(x = hidden_states_11_cast_fp16, y = attn_out_35)[name = string("x_197_cast_fp16")];
            fp16 var_5_promoted_23_to_fp16 = const()[name = string("op_5_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_1246_cast_fp16 = pow(x = x_197_cast_fp16, y = var_5_promoted_23_to_fp16)[name = string("op_1246_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1246_cast_fp16)[name = string("var_47_cast_fp16")];
            fp16 var_1249_to_fp16 = const()[name = string("op_1249_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1250_cast_fp16 = add(x = var_47_cast_fp16, y = var_1249_to_fp16)[name = string("op_1250_cast_fp16")];
            fp32 var_1251_epsilon_0 = const()[name = string("op_1251_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1251_cast_fp16 = rsqrt(epsilon = var_1251_epsilon_0, x = var_1250_cast_fp16)[name = string("op_1251_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_201_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1251_cast_fp16)[name = string("x_201_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_5_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_5_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1116523840)))];
            tensor<fp16, [1, 1024, 1024]> var_1254_cast_fp16 = mul(x = x_201_cast_fp16, y = encoder_layers_5_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_1254_cast_fp16")];
            tensor<int32, [3]> var_1259 = const()[name = string("op_1259"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_55_axes_0 = const()[name = string("input_55_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1260 = transpose(perm = var_1259, x = var_1254_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [1, 1024, 1, 1024]> input_55 = expand_dims(axes = input_55_axes_0, x = var_1260)[name = string("input_55")];
            string input_57_pad_type_0 = const()[name = string("input_57_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_57_strides_0 = const()[name = string("input_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_57_pad_0 = const()[name = string("input_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_57_dilations_0 = const()[name = string("input_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_57_groups_0 = const()[name = string("input_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_57 = conv(dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = encoder_layers_5_mlp_gate_proj_weight, x = input_55)[name = string("input_57")];
            string up_11_pad_type_0 = const()[name = string("up_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_11_strides_0 = const()[name = string("up_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_11_pad_0 = const()[name = string("up_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_11_dilations_0 = const()[name = string("up_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_11_groups_0 = const()[name = string("up_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_11 = conv(dilations = up_11_dilations_0, groups = up_11_groups_0, pad = up_11_pad_0, pad_type = up_11_pad_type_0, strides = up_11_strides_0, weight = encoder_layers_5_mlp_up_proj_weight, x = input_55)[name = string("up_11")];
            tensor<fp16, [1, 3072, 1, 1024]> var_1274 = silu(x = input_57)[name = string("op_1274")];
            tensor<fp16, [1, 3072, 1, 1024]> input_59 = mul(x = var_1274, y = up_11)[name = string("input_59")];
            string var_1281_pad_type_0 = const()[name = string("op_1281_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1281_strides_0 = const()[name = string("op_1281_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1281_pad_0 = const()[name = string("op_1281_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1281_dilations_0 = const()[name = string("op_1281_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1281_groups_0 = const()[name = string("op_1281_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1281 = conv(dilations = var_1281_dilations_0, groups = var_1281_groups_0, pad = var_1281_pad_0, pad_type = var_1281_pad_type_0, strides = var_1281_strides_0, weight = encoder_layers_5_mlp_down_proj_weight, x = input_59)[name = string("op_1281")];
            tensor<int32, [1]> var_1282_axes_0 = const()[name = string("op_1282_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1282 = squeeze(axes = var_1282_axes_0, x = var_1281)[name = string("op_1282")];
            tensor<int32, [3]> var_1283 = const()[name = string("op_1283"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_11 = transpose(perm = var_1283, x = var_1282)[name = string("transpose_198")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_13_cast_fp16 = add(x = x_197_cast_fp16, y = mlp_out_11)[name = string("hidden_states_13_cast_fp16")];
            fp16 var_5_promoted_24_to_fp16 = const()[name = string("op_5_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_1310_cast_fp16 = pow(x = hidden_states_13_cast_fp16, y = var_5_promoted_24_to_fp16)[name = string("op_1310_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1310_cast_fp16)[name = string("var_49_cast_fp16")];
            fp16 var_1313_to_fp16 = const()[name = string("op_1313_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1314_cast_fp16 = add(x = var_49_cast_fp16, y = var_1313_to_fp16)[name = string("op_1314_cast_fp16")];
            fp32 var_1315_epsilon_0 = const()[name = string("op_1315_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1315_cast_fp16 = rsqrt(epsilon = var_1315_epsilon_0, x = var_1314_cast_fp16)[name = string("op_1315_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_207_cast_fp16 = mul(x = hidden_states_13_cast_fp16, y = var_1315_cast_fp16)[name = string("x_207_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_6_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_6_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1116525952)))];
            tensor<fp16, [1, 1024, 1024]> var_1318_cast_fp16 = mul(x = x_207_cast_fp16, y = encoder_layers_6_input_layernorm_weight_promoted_to_fp16)[name = string("op_1318_cast_fp16")];
            tensor<int32, [3]> var_1323 = const()[name = string("op_1323"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_61_axes_0 = const()[name = string("input_61_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1324 = transpose(perm = var_1323, x = var_1318_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 1024, 1, 1024]> input_61 = expand_dims(axes = input_61_axes_0, x = var_1324)[name = string("input_61")];
            string var_1331_pad_type_0 = const()[name = string("op_1331_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1331_strides_0 = const()[name = string("op_1331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1331_pad_0 = const()[name = string("op_1331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1331_dilations_0 = const()[name = string("op_1331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1331_groups_0 = const()[name = string("op_1331_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_1331 = conv(dilations = var_1331_dilations_0, groups = var_1331_groups_0, pad = var_1331_pad_0, pad_type = var_1331_pad_type_0, strides = var_1331_strides_0, weight = encoder_layers_6_self_attn_q_proj_weight, x = input_61)[name = string("op_1331")];
            tensor<int32, [4]> var_1332 = const()[name = string("op_1332"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_1333 = reshape(shape = var_1332, x = var_1331)[name = string("op_1333")];
            tensor<int32, [4]> var_1334 = const()[name = string("op_1334"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1341_pad_type_0 = const()[name = string("op_1341_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1341_strides_0 = const()[name = string("op_1341_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1341_pad_0 = const()[name = string("op_1341_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1341_dilations_0 = const()[name = string("op_1341_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1341_groups_0 = const()[name = string("op_1341_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1341 = conv(dilations = var_1341_dilations_0, groups = var_1341_groups_0, pad = var_1341_pad_0, pad_type = var_1341_pad_type_0, strides = var_1341_strides_0, weight = encoder_layers_6_self_attn_k_proj_weight, x = input_61)[name = string("op_1341")];
            tensor<int32, [4]> var_1342 = const()[name = string("op_1342"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_1343 = reshape(shape = var_1342, x = var_1341)[name = string("op_1343")];
            tensor<int32, [4]> var_1344 = const()[name = string("op_1344"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1351_pad_type_0 = const()[name = string("op_1351_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1351_strides_0 = const()[name = string("op_1351_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1351_pad_0 = const()[name = string("op_1351_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1351_dilations_0 = const()[name = string("op_1351_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1351_groups_0 = const()[name = string("op_1351_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1351 = conv(dilations = var_1351_dilations_0, groups = var_1351_groups_0, pad = var_1351_pad_0, pad_type = var_1351_pad_type_0, strides = var_1351_strides_0, weight = encoder_layers_6_self_attn_v_proj_weight, x = input_61)[name = string("op_1351")];
            tensor<int32, [4]> var_1352 = const()[name = string("op_1352"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_1353 = reshape(shape = var_1352, x = var_1351)[name = string("op_1353")];
            tensor<int32, [4]> var_1354 = const()[name = string("op_1354"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_25_to_fp16 = const()[name = string("op_5_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_37 = transpose(perm = var_1334, x = var_1333)[name = string("transpose_196")];
            tensor<fp16, [1, 16, 1024, 128]> var_1360_cast_fp16 = pow(x = q_37, y = var_5_promoted_25_to_fp16)[name = string("op_1360_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_1360_cast_fp16)[name = string("var_51_cast_fp16")];
            fp16 var_1363_to_fp16 = const()[name = string("op_1363_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_1364_cast_fp16 = add(x = var_51_cast_fp16, y = var_1363_to_fp16)[name = string("op_1364_cast_fp16")];
            fp32 var_1365_epsilon_0 = const()[name = string("op_1365_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_1365_cast_fp16 = rsqrt(epsilon = var_1365_epsilon_0, x = var_1364_cast_fp16)[name = string("op_1365_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_215_cast_fp16 = mul(x = q_37, y = var_1365_cast_fp16)[name = string("x_215_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_39 = mul(x = x_215_cast_fp16, y = encoder_layers_6_self_attn_q_norm_weight)[name = string("q_39")];
            fp16 var_5_promoted_26_to_fp16 = const()[name = string("op_5_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_37 = transpose(perm = var_1344, x = var_1343)[name = string("transpose_195")];
            tensor<fp16, [1, 8, 1024, 128]> var_1373_cast_fp16 = pow(x = k_37, y = var_5_promoted_26_to_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_1373_cast_fp16)[name = string("var_53_cast_fp16")];
            fp16 var_1376_to_fp16 = const()[name = string("op_1376_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_1377_cast_fp16 = add(x = var_53_cast_fp16, y = var_1376_to_fp16)[name = string("op_1377_cast_fp16")];
            fp32 var_1378_epsilon_0 = const()[name = string("op_1378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_1378_cast_fp16 = rsqrt(epsilon = var_1378_epsilon_0, x = var_1377_cast_fp16)[name = string("op_1378_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_221_cast_fp16 = mul(x = k_37, y = var_1378_cast_fp16)[name = string("x_221_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_39 = mul(x = x_221_cast_fp16, y = encoder_layers_6_self_attn_k_norm_weight)[name = string("k_39")];
            tensor<fp16, [1, 16, 1024, 128]> var_1382 = mul(x = q_39, y = cos)[name = string("op_1382")];
            tensor<int32, [2]> var_1383_split_sizes_0 = const()[name = string("op_1383_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1383_axis_0 = const()[name = string("op_1383_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_1383_0, tensor<fp16, [1, 16, 1024, 64]> var_1383_1 = split(axis = var_1383_axis_0, split_sizes = var_1383_split_sizes_0, x = q_39)[name = string("op_1383")];
            fp16 const_21_promoted = const()[name = string("const_21_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_1385 = mul(x = var_1383_1, y = const_21_promoted)[name = string("op_1385")];
            bool var_1387_interleave_0 = const()[name = string("op_1387_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_1387 = concat(axis = var_17, interleave = var_1387_interleave_0, values = (var_1385, var_1383_0))[name = string("op_1387")];
            tensor<fp16, [1, 16, 1024, 128]> var_1388 = mul(x = var_1387, y = sin)[name = string("op_1388")];
            tensor<fp16, [1, 16, 1024, 128]> query_13 = add(x = var_1382, y = var_1388)[name = string("query_13")];
            tensor<fp16, [1, 8, 1024, 128]> var_1390 = mul(x = k_39, y = cos)[name = string("op_1390")];
            tensor<int32, [2]> var_1391_split_sizes_0 = const()[name = string("op_1391_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1391_axis_0 = const()[name = string("op_1391_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_1391_0, tensor<fp16, [1, 8, 1024, 64]> var_1391_1 = split(axis = var_1391_axis_0, split_sizes = var_1391_split_sizes_0, x = k_39)[name = string("op_1391")];
            fp16 const_22_promoted = const()[name = string("const_22_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_1393 = mul(x = var_1391_1, y = const_22_promoted)[name = string("op_1393")];
            bool var_1395_interleave_0 = const()[name = string("op_1395_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_1395 = concat(axis = var_17, interleave = var_1395_interleave_0, values = (var_1393, var_1391_0))[name = string("op_1395")];
            tensor<fp16, [1, 8, 1024, 128]> var_1396 = mul(x = var_1395, y = sin)[name = string("op_1396")];
            tensor<fp16, [1, 8, 1024, 128]> x_223 = add(x = var_1390, y = var_1396)[name = string("x_223")];
            tensor<int32, [1]> var_1398_axes_0 = const()[name = string("op_1398_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1398 = expand_dims(axes = var_1398_axes_0, x = x_223)[name = string("op_1398")];
            tensor<int32, [5]> x_225_reps_0 = const()[name = string("x_225_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_225 = tile(reps = x_225_reps_0, x = var_1398)[name = string("x_225")];
            tensor<int32, [4]> var_1401 = const()[name = string("op_1401"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_13 = reshape(shape = var_1401, x = x_225)[name = string("key_13")];
            tensor<int32, [1]> var_1403_axes_0 = const()[name = string("op_1403_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_227 = transpose(perm = var_1354, x = var_1353)[name = string("transpose_194")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1403 = expand_dims(axes = var_1403_axes_0, x = x_227)[name = string("op_1403")];
            tensor<int32, [5]> x_229_reps_0 = const()[name = string("x_229_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_229 = tile(reps = x_229_reps_0, x = var_1403)[name = string("x_229")];
            tensor<int32, [4]> var_1406 = const()[name = string("op_1406"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_13 = reshape(shape = var_1406, x = x_229)[name = string("value_13")];
            bool var_1411_transpose_x_1 = const()[name = string("op_1411_transpose_x_1"), val = bool(false)];
            bool var_1411_transpose_y_1 = const()[name = string("op_1411_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_1411_cast_fp16 = matmul(transpose_x = var_1411_transpose_x_1, transpose_y = var_1411_transpose_y_1, x = query_13, y = key_13)[name = string("op_1411_cast_fp16")];
            fp16 var_1412_to_fp16 = const()[name = string("op_1412_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_37_cast_fp16 = mul(x = var_1411_cast_fp16, y = var_1412_to_fp16)[name = string("attn_weights_37_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_39_cast_fp16 = add(x = attn_weights_37_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1416_cast_fp16 = softmax(axis = var_17, x = attn_weights_39_cast_fp16)[name = string("op_1416_cast_fp16")];
            bool var_1420_transpose_x_0 = const()[name = string("op_1420_transpose_x_0"), val = bool(false)];
            bool var_1420_transpose_y_0 = const()[name = string("op_1420_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_1420_cast_fp16 = matmul(transpose_x = var_1420_transpose_x_0, transpose_y = var_1420_transpose_y_0, x = var_1416_cast_fp16, y = value_13)[name = string("op_1420_cast_fp16")];
            tensor<int32, [4]> var_1422 = const()[name = string("op_1422"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1425 = const()[name = string("op_1425"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_1423 = transpose(perm = var_1422, x = var_1420_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [1, 1024, 2048]> attn_out_39 = reshape(shape = var_1425, x = var_1423)[name = string("attn_out_39")];
            tensor<int32, [3]> var_1427 = const()[name = string("op_1427"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_6 = const()[name = string("squeeze_6"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1116528064)))];
            string var_1436_pad_type_0 = const()[name = string("op_1436_pad_type_0"), val = string("valid")];
            int32 var_1436_groups_0 = const()[name = string("op_1436_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_1436_strides_0 = const()[name = string("op_1436_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1436_pad_0 = const()[name = string("op_1436_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1436_dilations_0 = const()[name = string("op_1436_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_1428 = transpose(perm = var_1427, x = attn_out_39)[name = string("transpose_192")];
            tensor<fp16, [1, 1024, 1024]> var_1436 = conv(dilations = var_1436_dilations_0, groups = var_1436_groups_0, pad = var_1436_pad_0, pad_type = var_1436_pad_type_0, strides = var_1436_strides_0, weight = squeeze_6, x = var_1428)[name = string("op_1436")];
            tensor<int32, [3]> var_1437 = const()[name = string("op_1437"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_41 = transpose(perm = var_1437, x = var_1436)[name = string("transpose_191")];
            tensor<fp16, [1, 1024, 1024]> x_231_cast_fp16 = add(x = hidden_states_13_cast_fp16, y = attn_out_41)[name = string("x_231_cast_fp16")];
            fp16 var_5_promoted_27_to_fp16 = const()[name = string("op_5_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_1443_cast_fp16 = pow(x = x_231_cast_fp16, y = var_5_promoted_27_to_fp16)[name = string("op_1443_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_1443_cast_fp16)[name = string("var_55_cast_fp16")];
            fp16 var_1446_to_fp16 = const()[name = string("op_1446_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1447_cast_fp16 = add(x = var_55_cast_fp16, y = var_1446_to_fp16)[name = string("op_1447_cast_fp16")];
            fp32 var_1448_epsilon_0 = const()[name = string("op_1448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1448_cast_fp16 = rsqrt(epsilon = var_1448_epsilon_0, x = var_1447_cast_fp16)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_235_cast_fp16 = mul(x = x_231_cast_fp16, y = var_1448_cast_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_6_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_6_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1120722432)))];
            tensor<fp16, [1, 1024, 1024]> var_1451_cast_fp16 = mul(x = x_235_cast_fp16, y = encoder_layers_6_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [3]> var_1456 = const()[name = string("op_1456"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1457 = transpose(perm = var_1456, x = var_1451_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [1, 1024, 1, 1024]> input_65 = expand_dims(axes = input_65_axes_0, x = var_1457)[name = string("input_65")];
            string input_67_pad_type_0 = const()[name = string("input_67_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_67_strides_0 = const()[name = string("input_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_67_pad_0 = const()[name = string("input_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_67_dilations_0 = const()[name = string("input_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_67_groups_0 = const()[name = string("input_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_67 = conv(dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = encoder_layers_6_mlp_gate_proj_weight, x = input_65)[name = string("input_67")];
            string up_13_pad_type_0 = const()[name = string("up_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_13_strides_0 = const()[name = string("up_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_13_pad_0 = const()[name = string("up_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_13_dilations_0 = const()[name = string("up_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_13_groups_0 = const()[name = string("up_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_13 = conv(dilations = up_13_dilations_0, groups = up_13_groups_0, pad = up_13_pad_0, pad_type = up_13_pad_type_0, strides = up_13_strides_0, weight = encoder_layers_6_mlp_up_proj_weight, x = input_65)[name = string("up_13")];
            tensor<fp16, [1, 3072, 1, 1024]> var_1471 = silu(x = input_67)[name = string("op_1471")];
            tensor<fp16, [1, 3072, 1, 1024]> input_69 = mul(x = var_1471, y = up_13)[name = string("input_69")];
            string var_1478_pad_type_0 = const()[name = string("op_1478_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1478_strides_0 = const()[name = string("op_1478_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1478_pad_0 = const()[name = string("op_1478_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1478_dilations_0 = const()[name = string("op_1478_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1478_groups_0 = const()[name = string("op_1478_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1478 = conv(dilations = var_1478_dilations_0, groups = var_1478_groups_0, pad = var_1478_pad_0, pad_type = var_1478_pad_type_0, strides = var_1478_strides_0, weight = encoder_layers_6_mlp_down_proj_weight, x = input_69)[name = string("op_1478")];
            tensor<int32, [1]> var_1479_axes_0 = const()[name = string("op_1479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1479 = squeeze(axes = var_1479_axes_0, x = var_1478)[name = string("op_1479")];
            tensor<int32, [3]> var_1480 = const()[name = string("op_1480"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_13 = transpose(perm = var_1480, x = var_1479)[name = string("transpose_189")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_15_cast_fp16 = add(x = x_231_cast_fp16, y = mlp_out_13)[name = string("hidden_states_15_cast_fp16")];
            fp16 var_5_promoted_28_to_fp16 = const()[name = string("op_5_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_1507_cast_fp16 = pow(x = hidden_states_15_cast_fp16, y = var_5_promoted_28_to_fp16)[name = string("op_1507_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_1507_cast_fp16)[name = string("var_57_cast_fp16")];
            fp16 var_1510_to_fp16 = const()[name = string("op_1510_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1511_cast_fp16 = add(x = var_57_cast_fp16, y = var_1510_to_fp16)[name = string("op_1511_cast_fp16")];
            fp32 var_1512_epsilon_0 = const()[name = string("op_1512_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1512_cast_fp16 = rsqrt(epsilon = var_1512_epsilon_0, x = var_1511_cast_fp16)[name = string("op_1512_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_241_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = var_1512_cast_fp16)[name = string("x_241_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_7_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_7_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1120724544)))];
            tensor<fp16, [1, 1024, 1024]> var_1515_cast_fp16 = mul(x = x_241_cast_fp16, y = encoder_layers_7_input_layernorm_weight_promoted_to_fp16)[name = string("op_1515_cast_fp16")];
            tensor<int32, [3]> var_1520 = const()[name = string("op_1520"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_71_axes_0 = const()[name = string("input_71_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1521 = transpose(perm = var_1520, x = var_1515_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [1, 1024, 1, 1024]> input_71 = expand_dims(axes = input_71_axes_0, x = var_1521)[name = string("input_71")];
            string var_1528_pad_type_0 = const()[name = string("op_1528_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1528_strides_0 = const()[name = string("op_1528_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1528_pad_0 = const()[name = string("op_1528_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1528_dilations_0 = const()[name = string("op_1528_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1528_groups_0 = const()[name = string("op_1528_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_1528 = conv(dilations = var_1528_dilations_0, groups = var_1528_groups_0, pad = var_1528_pad_0, pad_type = var_1528_pad_type_0, strides = var_1528_strides_0, weight = encoder_layers_7_self_attn_q_proj_weight, x = input_71)[name = string("op_1528")];
            tensor<int32, [4]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_1530 = reshape(shape = var_1529, x = var_1528)[name = string("op_1530")];
            tensor<int32, [4]> var_1531 = const()[name = string("op_1531"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1538_pad_type_0 = const()[name = string("op_1538_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1538_strides_0 = const()[name = string("op_1538_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1538_pad_0 = const()[name = string("op_1538_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1538_dilations_0 = const()[name = string("op_1538_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1538_groups_0 = const()[name = string("op_1538_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1538 = conv(dilations = var_1538_dilations_0, groups = var_1538_groups_0, pad = var_1538_pad_0, pad_type = var_1538_pad_type_0, strides = var_1538_strides_0, weight = encoder_layers_7_self_attn_k_proj_weight, x = input_71)[name = string("op_1538")];
            tensor<int32, [4]> var_1539 = const()[name = string("op_1539"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_1540 = reshape(shape = var_1539, x = var_1538)[name = string("op_1540")];
            tensor<int32, [4]> var_1541 = const()[name = string("op_1541"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1548_pad_type_0 = const()[name = string("op_1548_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1548_strides_0 = const()[name = string("op_1548_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1548_pad_0 = const()[name = string("op_1548_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1548_dilations_0 = const()[name = string("op_1548_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1548_groups_0 = const()[name = string("op_1548_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1548 = conv(dilations = var_1548_dilations_0, groups = var_1548_groups_0, pad = var_1548_pad_0, pad_type = var_1548_pad_type_0, strides = var_1548_strides_0, weight = encoder_layers_7_self_attn_v_proj_weight, x = input_71)[name = string("op_1548")];
            tensor<int32, [4]> var_1549 = const()[name = string("op_1549"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_1550 = reshape(shape = var_1549, x = var_1548)[name = string("op_1550")];
            tensor<int32, [4]> var_1551 = const()[name = string("op_1551"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_29_to_fp16 = const()[name = string("op_5_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_43 = transpose(perm = var_1531, x = var_1530)[name = string("transpose_187")];
            tensor<fp16, [1, 16, 1024, 128]> var_1557_cast_fp16 = pow(x = q_43, y = var_5_promoted_29_to_fp16)[name = string("op_1557_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_1557_cast_fp16)[name = string("var_59_cast_fp16")];
            fp16 var_1560_to_fp16 = const()[name = string("op_1560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_1561_cast_fp16 = add(x = var_59_cast_fp16, y = var_1560_to_fp16)[name = string("op_1561_cast_fp16")];
            fp32 var_1562_epsilon_0 = const()[name = string("op_1562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_1562_cast_fp16 = rsqrt(epsilon = var_1562_epsilon_0, x = var_1561_cast_fp16)[name = string("op_1562_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_249_cast_fp16 = mul(x = q_43, y = var_1562_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_45 = mul(x = x_249_cast_fp16, y = encoder_layers_7_self_attn_q_norm_weight)[name = string("q_45")];
            fp16 var_5_promoted_30_to_fp16 = const()[name = string("op_5_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_43 = transpose(perm = var_1541, x = var_1540)[name = string("transpose_186")];
            tensor<fp16, [1, 8, 1024, 128]> var_1570_cast_fp16 = pow(x = k_43, y = var_5_promoted_30_to_fp16)[name = string("op_1570_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_1570_cast_fp16)[name = string("var_61_cast_fp16")];
            fp16 var_1573_to_fp16 = const()[name = string("op_1573_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_1574_cast_fp16 = add(x = var_61_cast_fp16, y = var_1573_to_fp16)[name = string("op_1574_cast_fp16")];
            fp32 var_1575_epsilon_0 = const()[name = string("op_1575_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_1575_cast_fp16 = rsqrt(epsilon = var_1575_epsilon_0, x = var_1574_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_255_cast_fp16 = mul(x = k_43, y = var_1575_cast_fp16)[name = string("x_255_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_45 = mul(x = x_255_cast_fp16, y = encoder_layers_7_self_attn_k_norm_weight)[name = string("k_45")];
            tensor<fp16, [1, 16, 1024, 128]> var_1579 = mul(x = q_45, y = cos)[name = string("op_1579")];
            tensor<int32, [2]> var_1580_split_sizes_0 = const()[name = string("op_1580_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1580_axis_0 = const()[name = string("op_1580_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_1580_0, tensor<fp16, [1, 16, 1024, 64]> var_1580_1 = split(axis = var_1580_axis_0, split_sizes = var_1580_split_sizes_0, x = q_45)[name = string("op_1580")];
            fp16 const_24_promoted = const()[name = string("const_24_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_1582 = mul(x = var_1580_1, y = const_24_promoted)[name = string("op_1582")];
            bool var_1584_interleave_0 = const()[name = string("op_1584_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_1584 = concat(axis = var_17, interleave = var_1584_interleave_0, values = (var_1582, var_1580_0))[name = string("op_1584")];
            tensor<fp16, [1, 16, 1024, 128]> var_1585 = mul(x = var_1584, y = sin)[name = string("op_1585")];
            tensor<fp16, [1, 16, 1024, 128]> query_15 = add(x = var_1579, y = var_1585)[name = string("query_15")];
            tensor<fp16, [1, 8, 1024, 128]> var_1587 = mul(x = k_45, y = cos)[name = string("op_1587")];
            tensor<int32, [2]> var_1588_split_sizes_0 = const()[name = string("op_1588_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1588_axis_0 = const()[name = string("op_1588_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_1588_0, tensor<fp16, [1, 8, 1024, 64]> var_1588_1 = split(axis = var_1588_axis_0, split_sizes = var_1588_split_sizes_0, x = k_45)[name = string("op_1588")];
            fp16 const_25_promoted = const()[name = string("const_25_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_1590 = mul(x = var_1588_1, y = const_25_promoted)[name = string("op_1590")];
            bool var_1592_interleave_0 = const()[name = string("op_1592_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_1592 = concat(axis = var_17, interleave = var_1592_interleave_0, values = (var_1590, var_1588_0))[name = string("op_1592")];
            tensor<fp16, [1, 8, 1024, 128]> var_1593 = mul(x = var_1592, y = sin)[name = string("op_1593")];
            tensor<fp16, [1, 8, 1024, 128]> x_257 = add(x = var_1587, y = var_1593)[name = string("x_257")];
            tensor<int32, [1]> var_1595_axes_0 = const()[name = string("op_1595_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1595 = expand_dims(axes = var_1595_axes_0, x = x_257)[name = string("op_1595")];
            tensor<int32, [5]> x_259_reps_0 = const()[name = string("x_259_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_259 = tile(reps = x_259_reps_0, x = var_1595)[name = string("x_259")];
            tensor<int32, [4]> var_1598 = const()[name = string("op_1598"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_15 = reshape(shape = var_1598, x = x_259)[name = string("key_15")];
            tensor<int32, [1]> var_1600_axes_0 = const()[name = string("op_1600_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_261 = transpose(perm = var_1551, x = var_1550)[name = string("transpose_185")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1600 = expand_dims(axes = var_1600_axes_0, x = x_261)[name = string("op_1600")];
            tensor<int32, [5]> x_263_reps_0 = const()[name = string("x_263_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_263 = tile(reps = x_263_reps_0, x = var_1600)[name = string("x_263")];
            tensor<int32, [4]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_15 = reshape(shape = var_1603, x = x_263)[name = string("value_15")];
            bool var_1608_transpose_x_1 = const()[name = string("op_1608_transpose_x_1"), val = bool(false)];
            bool var_1608_transpose_y_1 = const()[name = string("op_1608_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_1608_cast_fp16 = matmul(transpose_x = var_1608_transpose_x_1, transpose_y = var_1608_transpose_y_1, x = query_15, y = key_15)[name = string("op_1608_cast_fp16")];
            fp16 var_1609_to_fp16 = const()[name = string("op_1609_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_43_cast_fp16 = mul(x = var_1608_cast_fp16, y = var_1609_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1613_cast_fp16 = softmax(axis = var_17, x = attn_weights_45_cast_fp16)[name = string("op_1613_cast_fp16")];
            bool var_1617_transpose_x_0 = const()[name = string("op_1617_transpose_x_0"), val = bool(false)];
            bool var_1617_transpose_y_0 = const()[name = string("op_1617_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_1617_cast_fp16 = matmul(transpose_x = var_1617_transpose_x_0, transpose_y = var_1617_transpose_y_0, x = var_1613_cast_fp16, y = value_15)[name = string("op_1617_cast_fp16")];
            tensor<int32, [4]> var_1619 = const()[name = string("op_1619"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1622 = const()[name = string("op_1622"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_1620 = transpose(perm = var_1619, x = var_1617_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [1, 1024, 2048]> attn_out_45 = reshape(shape = var_1622, x = var_1620)[name = string("attn_out_45")];
            tensor<int32, [3]> var_1624 = const()[name = string("op_1624"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_7 = const()[name = string("squeeze_7"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1120726656)))];
            string var_1633_pad_type_0 = const()[name = string("op_1633_pad_type_0"), val = string("valid")];
            int32 var_1633_groups_0 = const()[name = string("op_1633_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_1633_strides_0 = const()[name = string("op_1633_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1633_pad_0 = const()[name = string("op_1633_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1633_dilations_0 = const()[name = string("op_1633_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_1625 = transpose(perm = var_1624, x = attn_out_45)[name = string("transpose_183")];
            tensor<fp16, [1, 1024, 1024]> var_1633 = conv(dilations = var_1633_dilations_0, groups = var_1633_groups_0, pad = var_1633_pad_0, pad_type = var_1633_pad_type_0, strides = var_1633_strides_0, weight = squeeze_7, x = var_1625)[name = string("op_1633")];
            tensor<int32, [3]> var_1634 = const()[name = string("op_1634"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_47 = transpose(perm = var_1634, x = var_1633)[name = string("transpose_182")];
            tensor<fp16, [1, 1024, 1024]> x_265_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = attn_out_47)[name = string("x_265_cast_fp16")];
            fp16 var_5_promoted_31_to_fp16 = const()[name = string("op_5_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_1640_cast_fp16 = pow(x = x_265_cast_fp16, y = var_5_promoted_31_to_fp16)[name = string("op_1640_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_1640_cast_fp16)[name = string("var_63_cast_fp16")];
            fp16 var_1643_to_fp16 = const()[name = string("op_1643_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1644_cast_fp16 = add(x = var_63_cast_fp16, y = var_1643_to_fp16)[name = string("op_1644_cast_fp16")];
            fp32 var_1645_epsilon_0 = const()[name = string("op_1645_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1645_cast_fp16 = rsqrt(epsilon = var_1645_epsilon_0, x = var_1644_cast_fp16)[name = string("op_1645_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_269_cast_fp16 = mul(x = x_265_cast_fp16, y = var_1645_cast_fp16)[name = string("x_269_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_7_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_7_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1124921024)))];
            tensor<fp16, [1, 1024, 1024]> var_1648_cast_fp16 = mul(x = x_269_cast_fp16, y = encoder_layers_7_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [3]> var_1653 = const()[name = string("op_1653"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_75_axes_0 = const()[name = string("input_75_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1654 = transpose(perm = var_1653, x = var_1648_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [1, 1024, 1, 1024]> input_75 = expand_dims(axes = input_75_axes_0, x = var_1654)[name = string("input_75")];
            string input_77_pad_type_0 = const()[name = string("input_77_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_77_strides_0 = const()[name = string("input_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_77_pad_0 = const()[name = string("input_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_77_dilations_0 = const()[name = string("input_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_77_groups_0 = const()[name = string("input_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_77 = conv(dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = encoder_layers_7_mlp_gate_proj_weight, x = input_75)[name = string("input_77")];
            string up_15_pad_type_0 = const()[name = string("up_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_15_strides_0 = const()[name = string("up_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_15_pad_0 = const()[name = string("up_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_15_dilations_0 = const()[name = string("up_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_15_groups_0 = const()[name = string("up_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_15 = conv(dilations = up_15_dilations_0, groups = up_15_groups_0, pad = up_15_pad_0, pad_type = up_15_pad_type_0, strides = up_15_strides_0, weight = encoder_layers_7_mlp_up_proj_weight, x = input_75)[name = string("up_15")];
            tensor<fp16, [1, 3072, 1, 1024]> var_1668 = silu(x = input_77)[name = string("op_1668")];
            tensor<fp16, [1, 3072, 1, 1024]> input_79 = mul(x = var_1668, y = up_15)[name = string("input_79")];
            string var_1675_pad_type_0 = const()[name = string("op_1675_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1675_strides_0 = const()[name = string("op_1675_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1675_pad_0 = const()[name = string("op_1675_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1675_dilations_0 = const()[name = string("op_1675_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1675_groups_0 = const()[name = string("op_1675_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1675 = conv(dilations = var_1675_dilations_0, groups = var_1675_groups_0, pad = var_1675_pad_0, pad_type = var_1675_pad_type_0, strides = var_1675_strides_0, weight = encoder_layers_7_mlp_down_proj_weight, x = input_79)[name = string("op_1675")];
            tensor<int32, [1]> var_1676_axes_0 = const()[name = string("op_1676_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1676 = squeeze(axes = var_1676_axes_0, x = var_1675)[name = string("op_1676")];
            tensor<int32, [3]> var_1677 = const()[name = string("op_1677"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_15 = transpose(perm = var_1677, x = var_1676)[name = string("transpose_180")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_17_cast_fp16 = add(x = x_265_cast_fp16, y = mlp_out_15)[name = string("hidden_states_17_cast_fp16")];
            fp16 var_5_promoted_32_to_fp16 = const()[name = string("op_5_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_1704_cast_fp16 = pow(x = hidden_states_17_cast_fp16, y = var_5_promoted_32_to_fp16)[name = string("op_1704_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_1704_cast_fp16)[name = string("var_65_cast_fp16")];
            fp16 var_1707_to_fp16 = const()[name = string("op_1707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1708_cast_fp16 = add(x = var_65_cast_fp16, y = var_1707_to_fp16)[name = string("op_1708_cast_fp16")];
            fp32 var_1709_epsilon_0 = const()[name = string("op_1709_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1709_cast_fp16 = rsqrt(epsilon = var_1709_epsilon_0, x = var_1708_cast_fp16)[name = string("op_1709_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_275_cast_fp16 = mul(x = hidden_states_17_cast_fp16, y = var_1709_cast_fp16)[name = string("x_275_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_8_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_8_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1124923136)))];
            tensor<fp16, [1, 1024, 1024]> var_1712_cast_fp16 = mul(x = x_275_cast_fp16, y = encoder_layers_8_input_layernorm_weight_promoted_to_fp16)[name = string("op_1712_cast_fp16")];
            tensor<int32, [3]> var_1717 = const()[name = string("op_1717"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_81_axes_0 = const()[name = string("input_81_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1718 = transpose(perm = var_1717, x = var_1712_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 1024, 1, 1024]> input_81 = expand_dims(axes = input_81_axes_0, x = var_1718)[name = string("input_81")];
            string var_1725_pad_type_0 = const()[name = string("op_1725_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1725_strides_0 = const()[name = string("op_1725_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1725_pad_0 = const()[name = string("op_1725_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1725_dilations_0 = const()[name = string("op_1725_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1725_groups_0 = const()[name = string("op_1725_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_1725 = conv(dilations = var_1725_dilations_0, groups = var_1725_groups_0, pad = var_1725_pad_0, pad_type = var_1725_pad_type_0, strides = var_1725_strides_0, weight = encoder_layers_8_self_attn_q_proj_weight, x = input_81)[name = string("op_1725")];
            tensor<int32, [4]> var_1726 = const()[name = string("op_1726"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_1727 = reshape(shape = var_1726, x = var_1725)[name = string("op_1727")];
            tensor<int32, [4]> var_1728 = const()[name = string("op_1728"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1735_pad_type_0 = const()[name = string("op_1735_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1735_strides_0 = const()[name = string("op_1735_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1735_pad_0 = const()[name = string("op_1735_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1735_dilations_0 = const()[name = string("op_1735_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1735_groups_0 = const()[name = string("op_1735_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1735 = conv(dilations = var_1735_dilations_0, groups = var_1735_groups_0, pad = var_1735_pad_0, pad_type = var_1735_pad_type_0, strides = var_1735_strides_0, weight = encoder_layers_8_self_attn_k_proj_weight, x = input_81)[name = string("op_1735")];
            tensor<int32, [4]> var_1736 = const()[name = string("op_1736"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_1737 = reshape(shape = var_1736, x = var_1735)[name = string("op_1737")];
            tensor<int32, [4]> var_1738 = const()[name = string("op_1738"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1745_pad_type_0 = const()[name = string("op_1745_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1745_strides_0 = const()[name = string("op_1745_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1745_pad_0 = const()[name = string("op_1745_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1745_dilations_0 = const()[name = string("op_1745_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1745_groups_0 = const()[name = string("op_1745_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1745 = conv(dilations = var_1745_dilations_0, groups = var_1745_groups_0, pad = var_1745_pad_0, pad_type = var_1745_pad_type_0, strides = var_1745_strides_0, weight = encoder_layers_8_self_attn_v_proj_weight, x = input_81)[name = string("op_1745")];
            tensor<int32, [4]> var_1746 = const()[name = string("op_1746"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_1747 = reshape(shape = var_1746, x = var_1745)[name = string("op_1747")];
            tensor<int32, [4]> var_1748 = const()[name = string("op_1748"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_33_to_fp16 = const()[name = string("op_5_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_49 = transpose(perm = var_1728, x = var_1727)[name = string("transpose_178")];
            tensor<fp16, [1, 16, 1024, 128]> var_1754_cast_fp16 = pow(x = q_49, y = var_5_promoted_33_to_fp16)[name = string("op_1754_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_1754_cast_fp16)[name = string("var_67_cast_fp16")];
            fp16 var_1757_to_fp16 = const()[name = string("op_1757_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_1758_cast_fp16 = add(x = var_67_cast_fp16, y = var_1757_to_fp16)[name = string("op_1758_cast_fp16")];
            fp32 var_1759_epsilon_0 = const()[name = string("op_1759_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_1759_cast_fp16 = rsqrt(epsilon = var_1759_epsilon_0, x = var_1758_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_283_cast_fp16 = mul(x = q_49, y = var_1759_cast_fp16)[name = string("x_283_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_51 = mul(x = x_283_cast_fp16, y = encoder_layers_8_self_attn_q_norm_weight)[name = string("q_51")];
            fp16 var_5_promoted_34_to_fp16 = const()[name = string("op_5_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_49 = transpose(perm = var_1738, x = var_1737)[name = string("transpose_177")];
            tensor<fp16, [1, 8, 1024, 128]> var_1767_cast_fp16 = pow(x = k_49, y = var_5_promoted_34_to_fp16)[name = string("op_1767_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_1767_cast_fp16)[name = string("var_69_cast_fp16")];
            fp16 var_1770_to_fp16 = const()[name = string("op_1770_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_1771_cast_fp16 = add(x = var_69_cast_fp16, y = var_1770_to_fp16)[name = string("op_1771_cast_fp16")];
            fp32 var_1772_epsilon_0 = const()[name = string("op_1772_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_1772_cast_fp16 = rsqrt(epsilon = var_1772_epsilon_0, x = var_1771_cast_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_289_cast_fp16 = mul(x = k_49, y = var_1772_cast_fp16)[name = string("x_289_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_51 = mul(x = x_289_cast_fp16, y = encoder_layers_8_self_attn_k_norm_weight)[name = string("k_51")];
            tensor<fp16, [1, 16, 1024, 128]> var_1776 = mul(x = q_51, y = cos)[name = string("op_1776")];
            tensor<int32, [2]> var_1777_split_sizes_0 = const()[name = string("op_1777_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1777_axis_0 = const()[name = string("op_1777_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_1777_0, tensor<fp16, [1, 16, 1024, 64]> var_1777_1 = split(axis = var_1777_axis_0, split_sizes = var_1777_split_sizes_0, x = q_51)[name = string("op_1777")];
            fp16 const_27_promoted = const()[name = string("const_27_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_1779 = mul(x = var_1777_1, y = const_27_promoted)[name = string("op_1779")];
            bool var_1781_interleave_0 = const()[name = string("op_1781_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_1781 = concat(axis = var_17, interleave = var_1781_interleave_0, values = (var_1779, var_1777_0))[name = string("op_1781")];
            tensor<fp16, [1, 16, 1024, 128]> var_1782 = mul(x = var_1781, y = sin)[name = string("op_1782")];
            tensor<fp16, [1, 16, 1024, 128]> query_17 = add(x = var_1776, y = var_1782)[name = string("query_17")];
            tensor<fp16, [1, 8, 1024, 128]> var_1784 = mul(x = k_51, y = cos)[name = string("op_1784")];
            tensor<int32, [2]> var_1785_split_sizes_0 = const()[name = string("op_1785_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1785_axis_0 = const()[name = string("op_1785_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_1785_0, tensor<fp16, [1, 8, 1024, 64]> var_1785_1 = split(axis = var_1785_axis_0, split_sizes = var_1785_split_sizes_0, x = k_51)[name = string("op_1785")];
            fp16 const_28_promoted = const()[name = string("const_28_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_1787 = mul(x = var_1785_1, y = const_28_promoted)[name = string("op_1787")];
            bool var_1789_interleave_0 = const()[name = string("op_1789_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_1789 = concat(axis = var_17, interleave = var_1789_interleave_0, values = (var_1787, var_1785_0))[name = string("op_1789")];
            tensor<fp16, [1, 8, 1024, 128]> var_1790 = mul(x = var_1789, y = sin)[name = string("op_1790")];
            tensor<fp16, [1, 8, 1024, 128]> x_291 = add(x = var_1784, y = var_1790)[name = string("x_291")];
            tensor<int32, [1]> var_1792_axes_0 = const()[name = string("op_1792_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1792 = expand_dims(axes = var_1792_axes_0, x = x_291)[name = string("op_1792")];
            tensor<int32, [5]> x_293_reps_0 = const()[name = string("x_293_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_293 = tile(reps = x_293_reps_0, x = var_1792)[name = string("x_293")];
            tensor<int32, [4]> var_1795 = const()[name = string("op_1795"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_17 = reshape(shape = var_1795, x = x_293)[name = string("key_17")];
            tensor<int32, [1]> var_1797_axes_0 = const()[name = string("op_1797_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_295 = transpose(perm = var_1748, x = var_1747)[name = string("transpose_176")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1797 = expand_dims(axes = var_1797_axes_0, x = x_295)[name = string("op_1797")];
            tensor<int32, [5]> x_297_reps_0 = const()[name = string("x_297_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_297 = tile(reps = x_297_reps_0, x = var_1797)[name = string("x_297")];
            tensor<int32, [4]> var_1800 = const()[name = string("op_1800"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_17 = reshape(shape = var_1800, x = x_297)[name = string("value_17")];
            bool var_1805_transpose_x_1 = const()[name = string("op_1805_transpose_x_1"), val = bool(false)];
            bool var_1805_transpose_y_1 = const()[name = string("op_1805_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_1805_cast_fp16 = matmul(transpose_x = var_1805_transpose_x_1, transpose_y = var_1805_transpose_y_1, x = query_17, y = key_17)[name = string("op_1805_cast_fp16")];
            fp16 var_1806_to_fp16 = const()[name = string("op_1806_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_49_cast_fp16 = mul(x = var_1805_cast_fp16, y = var_1806_to_fp16)[name = string("attn_weights_49_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_51_cast_fp16 = add(x = attn_weights_49_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1810_cast_fp16 = softmax(axis = var_17, x = attn_weights_51_cast_fp16)[name = string("op_1810_cast_fp16")];
            bool var_1814_transpose_x_0 = const()[name = string("op_1814_transpose_x_0"), val = bool(false)];
            bool var_1814_transpose_y_0 = const()[name = string("op_1814_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_1814_cast_fp16 = matmul(transpose_x = var_1814_transpose_x_0, transpose_y = var_1814_transpose_y_0, x = var_1810_cast_fp16, y = value_17)[name = string("op_1814_cast_fp16")];
            tensor<int32, [4]> var_1816 = const()[name = string("op_1816"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_1817 = transpose(perm = var_1816, x = var_1814_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [1, 1024, 2048]> attn_out_51 = reshape(shape = var_1819, x = var_1817)[name = string("attn_out_51")];
            tensor<int32, [3]> var_1821 = const()[name = string("op_1821"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_8 = const()[name = string("squeeze_8"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1124925248)))];
            string var_1830_pad_type_0 = const()[name = string("op_1830_pad_type_0"), val = string("valid")];
            int32 var_1830_groups_0 = const()[name = string("op_1830_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_1830_strides_0 = const()[name = string("op_1830_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1830_pad_0 = const()[name = string("op_1830_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1830_dilations_0 = const()[name = string("op_1830_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_1822 = transpose(perm = var_1821, x = attn_out_51)[name = string("transpose_174")];
            tensor<fp16, [1, 1024, 1024]> var_1830 = conv(dilations = var_1830_dilations_0, groups = var_1830_groups_0, pad = var_1830_pad_0, pad_type = var_1830_pad_type_0, strides = var_1830_strides_0, weight = squeeze_8, x = var_1822)[name = string("op_1830")];
            tensor<int32, [3]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_53 = transpose(perm = var_1831, x = var_1830)[name = string("transpose_173")];
            tensor<fp16, [1, 1024, 1024]> x_299_cast_fp16 = add(x = hidden_states_17_cast_fp16, y = attn_out_53)[name = string("x_299_cast_fp16")];
            fp16 var_5_promoted_35_to_fp16 = const()[name = string("op_5_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_1837_cast_fp16 = pow(x = x_299_cast_fp16, y = var_5_promoted_35_to_fp16)[name = string("op_1837_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_1837_cast_fp16)[name = string("var_71_cast_fp16")];
            fp16 var_1840_to_fp16 = const()[name = string("op_1840_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1841_cast_fp16 = add(x = var_71_cast_fp16, y = var_1840_to_fp16)[name = string("op_1841_cast_fp16")];
            fp32 var_1842_epsilon_0 = const()[name = string("op_1842_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1842_cast_fp16 = rsqrt(epsilon = var_1842_epsilon_0, x = var_1841_cast_fp16)[name = string("op_1842_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_303_cast_fp16 = mul(x = x_299_cast_fp16, y = var_1842_cast_fp16)[name = string("x_303_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_8_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_8_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1129119616)))];
            tensor<fp16, [1, 1024, 1024]> var_1845_cast_fp16 = mul(x = x_303_cast_fp16, y = encoder_layers_8_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_1845_cast_fp16")];
            tensor<int32, [3]> var_1850 = const()[name = string("op_1850"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_85_axes_0 = const()[name = string("input_85_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1851 = transpose(perm = var_1850, x = var_1845_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 1024, 1, 1024]> input_85 = expand_dims(axes = input_85_axes_0, x = var_1851)[name = string("input_85")];
            string input_87_pad_type_0 = const()[name = string("input_87_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_87_strides_0 = const()[name = string("input_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_87_pad_0 = const()[name = string("input_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_87_dilations_0 = const()[name = string("input_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_87_groups_0 = const()[name = string("input_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_87 = conv(dilations = input_87_dilations_0, groups = input_87_groups_0, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = input_87_strides_0, weight = encoder_layers_8_mlp_gate_proj_weight, x = input_85)[name = string("input_87")];
            string up_17_pad_type_0 = const()[name = string("up_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_17_strides_0 = const()[name = string("up_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_17_pad_0 = const()[name = string("up_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_17_dilations_0 = const()[name = string("up_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_17_groups_0 = const()[name = string("up_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_17 = conv(dilations = up_17_dilations_0, groups = up_17_groups_0, pad = up_17_pad_0, pad_type = up_17_pad_type_0, strides = up_17_strides_0, weight = encoder_layers_8_mlp_up_proj_weight, x = input_85)[name = string("up_17")];
            tensor<fp16, [1, 3072, 1, 1024]> var_1865 = silu(x = input_87)[name = string("op_1865")];
            tensor<fp16, [1, 3072, 1, 1024]> input_89 = mul(x = var_1865, y = up_17)[name = string("input_89")];
            string var_1872_pad_type_0 = const()[name = string("op_1872_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1872_strides_0 = const()[name = string("op_1872_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1872_pad_0 = const()[name = string("op_1872_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1872_dilations_0 = const()[name = string("op_1872_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1872_groups_0 = const()[name = string("op_1872_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1872 = conv(dilations = var_1872_dilations_0, groups = var_1872_groups_0, pad = var_1872_pad_0, pad_type = var_1872_pad_type_0, strides = var_1872_strides_0, weight = encoder_layers_8_mlp_down_proj_weight, x = input_89)[name = string("op_1872")];
            tensor<int32, [1]> var_1873_axes_0 = const()[name = string("op_1873_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1873 = squeeze(axes = var_1873_axes_0, x = var_1872)[name = string("op_1873")];
            tensor<int32, [3]> var_1874 = const()[name = string("op_1874"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_17 = transpose(perm = var_1874, x = var_1873)[name = string("transpose_171")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_19_cast_fp16 = add(x = x_299_cast_fp16, y = mlp_out_17)[name = string("hidden_states_19_cast_fp16")];
            fp16 var_5_promoted_36_to_fp16 = const()[name = string("op_5_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_1901_cast_fp16 = pow(x = hidden_states_19_cast_fp16, y = var_5_promoted_36_to_fp16)[name = string("op_1901_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_1901_cast_fp16)[name = string("var_73_cast_fp16")];
            fp16 var_1904_to_fp16 = const()[name = string("op_1904_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1905_cast_fp16 = add(x = var_73_cast_fp16, y = var_1904_to_fp16)[name = string("op_1905_cast_fp16")];
            fp32 var_1906_epsilon_0 = const()[name = string("op_1906_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1906_cast_fp16 = rsqrt(epsilon = var_1906_epsilon_0, x = var_1905_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_309_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = var_1906_cast_fp16)[name = string("x_309_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_9_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_9_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1129121728)))];
            tensor<fp16, [1, 1024, 1024]> var_1909_cast_fp16 = mul(x = x_309_cast_fp16, y = encoder_layers_9_input_layernorm_weight_promoted_to_fp16)[name = string("op_1909_cast_fp16")];
            tensor<int32, [3]> var_1914 = const()[name = string("op_1914"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_91_axes_0 = const()[name = string("input_91_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1915 = transpose(perm = var_1914, x = var_1909_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 1024, 1, 1024]> input_91 = expand_dims(axes = input_91_axes_0, x = var_1915)[name = string("input_91")];
            string var_1922_pad_type_0 = const()[name = string("op_1922_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1922_strides_0 = const()[name = string("op_1922_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1922_pad_0 = const()[name = string("op_1922_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1922_dilations_0 = const()[name = string("op_1922_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1922_groups_0 = const()[name = string("op_1922_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_1922 = conv(dilations = var_1922_dilations_0, groups = var_1922_groups_0, pad = var_1922_pad_0, pad_type = var_1922_pad_type_0, strides = var_1922_strides_0, weight = encoder_layers_9_self_attn_q_proj_weight, x = input_91)[name = string("op_1922")];
            tensor<int32, [4]> var_1923 = const()[name = string("op_1923"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_1924 = reshape(shape = var_1923, x = var_1922)[name = string("op_1924")];
            tensor<int32, [4]> var_1925 = const()[name = string("op_1925"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1932_pad_type_0 = const()[name = string("op_1932_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1932_strides_0 = const()[name = string("op_1932_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1932_pad_0 = const()[name = string("op_1932_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1932_dilations_0 = const()[name = string("op_1932_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1932_groups_0 = const()[name = string("op_1932_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1932 = conv(dilations = var_1932_dilations_0, groups = var_1932_groups_0, pad = var_1932_pad_0, pad_type = var_1932_pad_type_0, strides = var_1932_strides_0, weight = encoder_layers_9_self_attn_k_proj_weight, x = input_91)[name = string("op_1932")];
            tensor<int32, [4]> var_1933 = const()[name = string("op_1933"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_1934 = reshape(shape = var_1933, x = var_1932)[name = string("op_1934")];
            tensor<int32, [4]> var_1935 = const()[name = string("op_1935"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1942_pad_type_0 = const()[name = string("op_1942_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1942_strides_0 = const()[name = string("op_1942_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1942_pad_0 = const()[name = string("op_1942_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1942_dilations_0 = const()[name = string("op_1942_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1942_groups_0 = const()[name = string("op_1942_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_1942 = conv(dilations = var_1942_dilations_0, groups = var_1942_groups_0, pad = var_1942_pad_0, pad_type = var_1942_pad_type_0, strides = var_1942_strides_0, weight = encoder_layers_9_self_attn_v_proj_weight, x = input_91)[name = string("op_1942")];
            tensor<int32, [4]> var_1943 = const()[name = string("op_1943"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_1944 = reshape(shape = var_1943, x = var_1942)[name = string("op_1944")];
            tensor<int32, [4]> var_1945 = const()[name = string("op_1945"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_37_to_fp16 = const()[name = string("op_5_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_55 = transpose(perm = var_1925, x = var_1924)[name = string("transpose_169")];
            tensor<fp16, [1, 16, 1024, 128]> var_1951_cast_fp16 = pow(x = q_55, y = var_5_promoted_37_to_fp16)[name = string("op_1951_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_1951_cast_fp16)[name = string("var_75_cast_fp16")];
            fp16 var_1954_to_fp16 = const()[name = string("op_1954_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_1955_cast_fp16 = add(x = var_75_cast_fp16, y = var_1954_to_fp16)[name = string("op_1955_cast_fp16")];
            fp32 var_1956_epsilon_0 = const()[name = string("op_1956_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_1956_cast_fp16 = rsqrt(epsilon = var_1956_epsilon_0, x = var_1955_cast_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_317_cast_fp16 = mul(x = q_55, y = var_1956_cast_fp16)[name = string("x_317_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_57 = mul(x = x_317_cast_fp16, y = encoder_layers_9_self_attn_q_norm_weight)[name = string("q_57")];
            fp16 var_5_promoted_38_to_fp16 = const()[name = string("op_5_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_55 = transpose(perm = var_1935, x = var_1934)[name = string("transpose_168")];
            tensor<fp16, [1, 8, 1024, 128]> var_1964_cast_fp16 = pow(x = k_55, y = var_5_promoted_38_to_fp16)[name = string("op_1964_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_1964_cast_fp16)[name = string("var_77_cast_fp16")];
            fp16 var_1967_to_fp16 = const()[name = string("op_1967_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_1968_cast_fp16 = add(x = var_77_cast_fp16, y = var_1967_to_fp16)[name = string("op_1968_cast_fp16")];
            fp32 var_1969_epsilon_0 = const()[name = string("op_1969_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_1969_cast_fp16 = rsqrt(epsilon = var_1969_epsilon_0, x = var_1968_cast_fp16)[name = string("op_1969_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_323_cast_fp16 = mul(x = k_55, y = var_1969_cast_fp16)[name = string("x_323_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_57 = mul(x = x_323_cast_fp16, y = encoder_layers_9_self_attn_k_norm_weight)[name = string("k_57")];
            tensor<fp16, [1, 16, 1024, 128]> var_1973 = mul(x = q_57, y = cos)[name = string("op_1973")];
            tensor<int32, [2]> var_1974_split_sizes_0 = const()[name = string("op_1974_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1974_axis_0 = const()[name = string("op_1974_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_1974_0, tensor<fp16, [1, 16, 1024, 64]> var_1974_1 = split(axis = var_1974_axis_0, split_sizes = var_1974_split_sizes_0, x = q_57)[name = string("op_1974")];
            fp16 const_30_promoted = const()[name = string("const_30_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_1976 = mul(x = var_1974_1, y = const_30_promoted)[name = string("op_1976")];
            bool var_1978_interleave_0 = const()[name = string("op_1978_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_1978 = concat(axis = var_17, interleave = var_1978_interleave_0, values = (var_1976, var_1974_0))[name = string("op_1978")];
            tensor<fp16, [1, 16, 1024, 128]> var_1979 = mul(x = var_1978, y = sin)[name = string("op_1979")];
            tensor<fp16, [1, 16, 1024, 128]> query_19 = add(x = var_1973, y = var_1979)[name = string("query_19")];
            tensor<fp16, [1, 8, 1024, 128]> var_1981 = mul(x = k_57, y = cos)[name = string("op_1981")];
            tensor<int32, [2]> var_1982_split_sizes_0 = const()[name = string("op_1982_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1982_axis_0 = const()[name = string("op_1982_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_1982_0, tensor<fp16, [1, 8, 1024, 64]> var_1982_1 = split(axis = var_1982_axis_0, split_sizes = var_1982_split_sizes_0, x = k_57)[name = string("op_1982")];
            fp16 const_31_promoted = const()[name = string("const_31_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_1984 = mul(x = var_1982_1, y = const_31_promoted)[name = string("op_1984")];
            bool var_1986_interleave_0 = const()[name = string("op_1986_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_1986 = concat(axis = var_17, interleave = var_1986_interleave_0, values = (var_1984, var_1982_0))[name = string("op_1986")];
            tensor<fp16, [1, 8, 1024, 128]> var_1987 = mul(x = var_1986, y = sin)[name = string("op_1987")];
            tensor<fp16, [1, 8, 1024, 128]> x_325 = add(x = var_1981, y = var_1987)[name = string("x_325")];
            tensor<int32, [1]> var_1989_axes_0 = const()[name = string("op_1989_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1989 = expand_dims(axes = var_1989_axes_0, x = x_325)[name = string("op_1989")];
            tensor<int32, [5]> x_327_reps_0 = const()[name = string("x_327_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_327 = tile(reps = x_327_reps_0, x = var_1989)[name = string("x_327")];
            tensor<int32, [4]> var_1992 = const()[name = string("op_1992"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_19 = reshape(shape = var_1992, x = x_327)[name = string("key_19")];
            tensor<int32, [1]> var_1994_axes_0 = const()[name = string("op_1994_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_329 = transpose(perm = var_1945, x = var_1944)[name = string("transpose_167")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1994 = expand_dims(axes = var_1994_axes_0, x = x_329)[name = string("op_1994")];
            tensor<int32, [5]> x_331_reps_0 = const()[name = string("x_331_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_331 = tile(reps = x_331_reps_0, x = var_1994)[name = string("x_331")];
            tensor<int32, [4]> var_1997 = const()[name = string("op_1997"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_19 = reshape(shape = var_1997, x = x_331)[name = string("value_19")];
            bool var_2002_transpose_x_1 = const()[name = string("op_2002_transpose_x_1"), val = bool(false)];
            bool var_2002_transpose_y_1 = const()[name = string("op_2002_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_2002_cast_fp16 = matmul(transpose_x = var_2002_transpose_x_1, transpose_y = var_2002_transpose_y_1, x = query_19, y = key_19)[name = string("op_2002_cast_fp16")];
            fp16 var_2003_to_fp16 = const()[name = string("op_2003_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_55_cast_fp16 = mul(x = var_2002_cast_fp16, y = var_2003_to_fp16)[name = string("attn_weights_55_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_57_cast_fp16 = add(x = attn_weights_55_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_57_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_2007_cast_fp16 = softmax(axis = var_17, x = attn_weights_57_cast_fp16)[name = string("op_2007_cast_fp16")];
            bool var_2011_transpose_x_0 = const()[name = string("op_2011_transpose_x_0"), val = bool(false)];
            bool var_2011_transpose_y_0 = const()[name = string("op_2011_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_2011_cast_fp16 = matmul(transpose_x = var_2011_transpose_x_0, transpose_y = var_2011_transpose_y_0, x = var_2007_cast_fp16, y = value_19)[name = string("op_2011_cast_fp16")];
            tensor<int32, [4]> var_2013 = const()[name = string("op_2013"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2016 = const()[name = string("op_2016"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_2014 = transpose(perm = var_2013, x = var_2011_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [1, 1024, 2048]> attn_out_57 = reshape(shape = var_2016, x = var_2014)[name = string("attn_out_57")];
            tensor<int32, [3]> var_2018 = const()[name = string("op_2018"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_9 = const()[name = string("squeeze_9"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1129123840)))];
            string var_2027_pad_type_0 = const()[name = string("op_2027_pad_type_0"), val = string("valid")];
            int32 var_2027_groups_0 = const()[name = string("op_2027_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_2027_strides_0 = const()[name = string("op_2027_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_2027_pad_0 = const()[name = string("op_2027_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_2027_dilations_0 = const()[name = string("op_2027_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_2019 = transpose(perm = var_2018, x = attn_out_57)[name = string("transpose_165")];
            tensor<fp16, [1, 1024, 1024]> var_2027 = conv(dilations = var_2027_dilations_0, groups = var_2027_groups_0, pad = var_2027_pad_0, pad_type = var_2027_pad_type_0, strides = var_2027_strides_0, weight = squeeze_9, x = var_2019)[name = string("op_2027")];
            tensor<int32, [3]> var_2028 = const()[name = string("op_2028"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_59 = transpose(perm = var_2028, x = var_2027)[name = string("transpose_164")];
            tensor<fp16, [1, 1024, 1024]> x_333_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = attn_out_59)[name = string("x_333_cast_fp16")];
            fp16 var_5_promoted_39_to_fp16 = const()[name = string("op_5_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_2034_cast_fp16 = pow(x = x_333_cast_fp16, y = var_5_promoted_39_to_fp16)[name = string("op_2034_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_2034_cast_fp16)[name = string("var_79_cast_fp16")];
            fp16 var_2037_to_fp16 = const()[name = string("op_2037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2038_cast_fp16 = add(x = var_79_cast_fp16, y = var_2037_to_fp16)[name = string("op_2038_cast_fp16")];
            fp32 var_2039_epsilon_0 = const()[name = string("op_2039_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2039_cast_fp16 = rsqrt(epsilon = var_2039_epsilon_0, x = var_2038_cast_fp16)[name = string("op_2039_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_337_cast_fp16 = mul(x = x_333_cast_fp16, y = var_2039_cast_fp16)[name = string("x_337_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_9_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_9_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1133318208)))];
            tensor<fp16, [1, 1024, 1024]> var_2042_cast_fp16 = mul(x = x_337_cast_fp16, y = encoder_layers_9_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_2042_cast_fp16")];
            tensor<int32, [3]> var_2047 = const()[name = string("op_2047"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_95_axes_0 = const()[name = string("input_95_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2048 = transpose(perm = var_2047, x = var_2042_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [1, 1024, 1, 1024]> input_95 = expand_dims(axes = input_95_axes_0, x = var_2048)[name = string("input_95")];
            string input_97_pad_type_0 = const()[name = string("input_97_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_97_strides_0 = const()[name = string("input_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_97_pad_0 = const()[name = string("input_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_97_dilations_0 = const()[name = string("input_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_97_groups_0 = const()[name = string("input_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_97 = conv(dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = encoder_layers_9_mlp_gate_proj_weight, x = input_95)[name = string("input_97")];
            string up_19_pad_type_0 = const()[name = string("up_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_19_strides_0 = const()[name = string("up_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_19_pad_0 = const()[name = string("up_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_19_dilations_0 = const()[name = string("up_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_19_groups_0 = const()[name = string("up_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_19 = conv(dilations = up_19_dilations_0, groups = up_19_groups_0, pad = up_19_pad_0, pad_type = up_19_pad_type_0, strides = up_19_strides_0, weight = encoder_layers_9_mlp_up_proj_weight, x = input_95)[name = string("up_19")];
            tensor<fp16, [1, 3072, 1, 1024]> var_2062 = silu(x = input_97)[name = string("op_2062")];
            tensor<fp16, [1, 3072, 1, 1024]> input_99 = mul(x = var_2062, y = up_19)[name = string("input_99")];
            string var_2069_pad_type_0 = const()[name = string("op_2069_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2069_strides_0 = const()[name = string("op_2069_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2069_pad_0 = const()[name = string("op_2069_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2069_dilations_0 = const()[name = string("op_2069_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2069_groups_0 = const()[name = string("op_2069_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2069 = conv(dilations = var_2069_dilations_0, groups = var_2069_groups_0, pad = var_2069_pad_0, pad_type = var_2069_pad_type_0, strides = var_2069_strides_0, weight = encoder_layers_9_mlp_down_proj_weight, x = input_99)[name = string("op_2069")];
            tensor<int32, [1]> var_2070_axes_0 = const()[name = string("op_2070_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2070 = squeeze(axes = var_2070_axes_0, x = var_2069)[name = string("op_2070")];
            tensor<int32, [3]> var_2071 = const()[name = string("op_2071"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_19 = transpose(perm = var_2071, x = var_2070)[name = string("transpose_162")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_21_cast_fp16 = add(x = x_333_cast_fp16, y = mlp_out_19)[name = string("hidden_states_21_cast_fp16")];
            fp16 var_5_promoted_40_to_fp16 = const()[name = string("op_5_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_2098_cast_fp16 = pow(x = hidden_states_21_cast_fp16, y = var_5_promoted_40_to_fp16)[name = string("op_2098_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_2098_cast_fp16)[name = string("var_81_cast_fp16")];
            fp16 var_2101_to_fp16 = const()[name = string("op_2101_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2102_cast_fp16 = add(x = var_81_cast_fp16, y = var_2101_to_fp16)[name = string("op_2102_cast_fp16")];
            fp32 var_2103_epsilon_0 = const()[name = string("op_2103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2103_cast_fp16 = rsqrt(epsilon = var_2103_epsilon_0, x = var_2102_cast_fp16)[name = string("op_2103_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_343_cast_fp16 = mul(x = hidden_states_21_cast_fp16, y = var_2103_cast_fp16)[name = string("x_343_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_10_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_10_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1133320320)))];
            tensor<fp16, [1, 1024, 1024]> var_2106_cast_fp16 = mul(x = x_343_cast_fp16, y = encoder_layers_10_input_layernorm_weight_promoted_to_fp16)[name = string("op_2106_cast_fp16")];
            tensor<int32, [3]> var_2111 = const()[name = string("op_2111"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2112 = transpose(perm = var_2111, x = var_2106_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 1024, 1, 1024]> input_101 = expand_dims(axes = input_101_axes_0, x = var_2112)[name = string("input_101")];
            string var_2119_pad_type_0 = const()[name = string("op_2119_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2119_strides_0 = const()[name = string("op_2119_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2119_pad_0 = const()[name = string("op_2119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2119_dilations_0 = const()[name = string("op_2119_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2119_groups_0 = const()[name = string("op_2119_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_2119 = conv(dilations = var_2119_dilations_0, groups = var_2119_groups_0, pad = var_2119_pad_0, pad_type = var_2119_pad_type_0, strides = var_2119_strides_0, weight = encoder_layers_10_self_attn_q_proj_weight, x = input_101)[name = string("op_2119")];
            tensor<int32, [4]> var_2120 = const()[name = string("op_2120"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_2121 = reshape(shape = var_2120, x = var_2119)[name = string("op_2121")];
            tensor<int32, [4]> var_2122 = const()[name = string("op_2122"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2129_pad_type_0 = const()[name = string("op_2129_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2129_strides_0 = const()[name = string("op_2129_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2129_pad_0 = const()[name = string("op_2129_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2129_dilations_0 = const()[name = string("op_2129_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2129_groups_0 = const()[name = string("op_2129_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2129 = conv(dilations = var_2129_dilations_0, groups = var_2129_groups_0, pad = var_2129_pad_0, pad_type = var_2129_pad_type_0, strides = var_2129_strides_0, weight = encoder_layers_10_self_attn_k_proj_weight, x = input_101)[name = string("op_2129")];
            tensor<int32, [4]> var_2130 = const()[name = string("op_2130"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_2131 = reshape(shape = var_2130, x = var_2129)[name = string("op_2131")];
            tensor<int32, [4]> var_2132 = const()[name = string("op_2132"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2139_pad_type_0 = const()[name = string("op_2139_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2139_strides_0 = const()[name = string("op_2139_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2139_pad_0 = const()[name = string("op_2139_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2139_dilations_0 = const()[name = string("op_2139_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2139_groups_0 = const()[name = string("op_2139_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2139 = conv(dilations = var_2139_dilations_0, groups = var_2139_groups_0, pad = var_2139_pad_0, pad_type = var_2139_pad_type_0, strides = var_2139_strides_0, weight = encoder_layers_10_self_attn_v_proj_weight, x = input_101)[name = string("op_2139")];
            tensor<int32, [4]> var_2140 = const()[name = string("op_2140"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_2141 = reshape(shape = var_2140, x = var_2139)[name = string("op_2141")];
            tensor<int32, [4]> var_2142 = const()[name = string("op_2142"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_41_to_fp16 = const()[name = string("op_5_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_61 = transpose(perm = var_2122, x = var_2121)[name = string("transpose_160")];
            tensor<fp16, [1, 16, 1024, 128]> var_2148_cast_fp16 = pow(x = q_61, y = var_5_promoted_41_to_fp16)[name = string("op_2148_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_2148_cast_fp16)[name = string("var_83_cast_fp16")];
            fp16 var_2151_to_fp16 = const()[name = string("op_2151_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_2152_cast_fp16 = add(x = var_83_cast_fp16, y = var_2151_to_fp16)[name = string("op_2152_cast_fp16")];
            fp32 var_2153_epsilon_0 = const()[name = string("op_2153_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_2153_cast_fp16 = rsqrt(epsilon = var_2153_epsilon_0, x = var_2152_cast_fp16)[name = string("op_2153_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_351_cast_fp16 = mul(x = q_61, y = var_2153_cast_fp16)[name = string("x_351_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_63 = mul(x = x_351_cast_fp16, y = encoder_layers_10_self_attn_q_norm_weight)[name = string("q_63")];
            fp16 var_5_promoted_42_to_fp16 = const()[name = string("op_5_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_61 = transpose(perm = var_2132, x = var_2131)[name = string("transpose_159")];
            tensor<fp16, [1, 8, 1024, 128]> var_2161_cast_fp16 = pow(x = k_61, y = var_5_promoted_42_to_fp16)[name = string("op_2161_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_85_cast_fp16_0 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_2161_cast_fp16)[name = string("var_85_cast_fp16")];
            fp16 var_2164_to_fp16 = const()[name = string("op_2164_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_2165_cast_fp16 = add(x = var_85_cast_fp16_0, y = var_2164_to_fp16)[name = string("op_2165_cast_fp16")];
            fp32 var_2166_epsilon_0 = const()[name = string("op_2166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_2166_cast_fp16 = rsqrt(epsilon = var_2166_epsilon_0, x = var_2165_cast_fp16)[name = string("op_2166_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_357_cast_fp16 = mul(x = k_61, y = var_2166_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_63 = mul(x = x_357_cast_fp16, y = encoder_layers_10_self_attn_k_norm_weight)[name = string("k_63")];
            tensor<fp16, [1, 16, 1024, 128]> var_2170 = mul(x = q_63, y = cos)[name = string("op_2170")];
            tensor<int32, [2]> var_2171_split_sizes_0 = const()[name = string("op_2171_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2171_axis_0 = const()[name = string("op_2171_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_2171_0, tensor<fp16, [1, 16, 1024, 64]> var_2171_1 = split(axis = var_2171_axis_0, split_sizes = var_2171_split_sizes_0, x = q_63)[name = string("op_2171")];
            fp16 const_33_promoted = const()[name = string("const_33_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_2173 = mul(x = var_2171_1, y = const_33_promoted)[name = string("op_2173")];
            bool var_2175_interleave_0 = const()[name = string("op_2175_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_2175 = concat(axis = var_17, interleave = var_2175_interleave_0, values = (var_2173, var_2171_0))[name = string("op_2175")];
            tensor<fp16, [1, 16, 1024, 128]> var_2176 = mul(x = var_2175, y = sin)[name = string("op_2176")];
            tensor<fp16, [1, 16, 1024, 128]> query_21 = add(x = var_2170, y = var_2176)[name = string("query_21")];
            tensor<fp16, [1, 8, 1024, 128]> var_2178 = mul(x = k_63, y = cos)[name = string("op_2178")];
            tensor<int32, [2]> var_2179_split_sizes_0 = const()[name = string("op_2179_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2179_axis_0 = const()[name = string("op_2179_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_2179_0, tensor<fp16, [1, 8, 1024, 64]> var_2179_1 = split(axis = var_2179_axis_0, split_sizes = var_2179_split_sizes_0, x = k_63)[name = string("op_2179")];
            fp16 const_34_promoted = const()[name = string("const_34_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_2181 = mul(x = var_2179_1, y = const_34_promoted)[name = string("op_2181")];
            bool var_2183_interleave_0 = const()[name = string("op_2183_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_2183 = concat(axis = var_17, interleave = var_2183_interleave_0, values = (var_2181, var_2179_0))[name = string("op_2183")];
            tensor<fp16, [1, 8, 1024, 128]> var_2184 = mul(x = var_2183, y = sin)[name = string("op_2184")];
            tensor<fp16, [1, 8, 1024, 128]> x_359 = add(x = var_2178, y = var_2184)[name = string("x_359")];
            tensor<int32, [1]> var_2186_axes_0 = const()[name = string("op_2186_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2186 = expand_dims(axes = var_2186_axes_0, x = x_359)[name = string("op_2186")];
            tensor<int32, [5]> x_361_reps_0 = const()[name = string("x_361_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_361 = tile(reps = x_361_reps_0, x = var_2186)[name = string("x_361")];
            tensor<int32, [4]> var_2189 = const()[name = string("op_2189"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_21 = reshape(shape = var_2189, x = x_361)[name = string("key_21")];
            tensor<int32, [1]> var_2191_axes_0 = const()[name = string("op_2191_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_363 = transpose(perm = var_2142, x = var_2141)[name = string("transpose_158")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2191 = expand_dims(axes = var_2191_axes_0, x = x_363)[name = string("op_2191")];
            tensor<int32, [5]> x_365_reps_0 = const()[name = string("x_365_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_365 = tile(reps = x_365_reps_0, x = var_2191)[name = string("x_365")];
            tensor<int32, [4]> var_2194 = const()[name = string("op_2194"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_21 = reshape(shape = var_2194, x = x_365)[name = string("value_21")];
            bool var_2199_transpose_x_1 = const()[name = string("op_2199_transpose_x_1"), val = bool(false)];
            bool var_2199_transpose_y_1 = const()[name = string("op_2199_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_1, transpose_y = var_2199_transpose_y_1, x = query_21, y = key_21)[name = string("op_2199_cast_fp16")];
            fp16 var_2200_to_fp16 = const()[name = string("op_2200_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_61_cast_fp16 = mul(x = var_2199_cast_fp16, y = var_2200_to_fp16)[name = string("attn_weights_61_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_63_cast_fp16 = add(x = attn_weights_61_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_2204_cast_fp16 = softmax(axis = var_17, x = attn_weights_63_cast_fp16)[name = string("op_2204_cast_fp16")];
            bool var_2208_transpose_x_0 = const()[name = string("op_2208_transpose_x_0"), val = bool(false)];
            bool var_2208_transpose_y_0 = const()[name = string("op_2208_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_2208_cast_fp16 = matmul(transpose_x = var_2208_transpose_x_0, transpose_y = var_2208_transpose_y_0, x = var_2204_cast_fp16, y = value_21)[name = string("op_2208_cast_fp16")];
            tensor<int32, [4]> var_2210 = const()[name = string("op_2210"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2213 = const()[name = string("op_2213"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_2211 = transpose(perm = var_2210, x = var_2208_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 1024, 2048]> attn_out_63 = reshape(shape = var_2213, x = var_2211)[name = string("attn_out_63")];
            tensor<int32, [3]> var_2215 = const()[name = string("op_2215"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_10 = const()[name = string("squeeze_10"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1133322432)))];
            string var_2224_pad_type_0 = const()[name = string("op_2224_pad_type_0"), val = string("valid")];
            int32 var_2224_groups_0 = const()[name = string("op_2224_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_2224_strides_0 = const()[name = string("op_2224_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_2224_pad_0 = const()[name = string("op_2224_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_2224_dilations_0 = const()[name = string("op_2224_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_2216 = transpose(perm = var_2215, x = attn_out_63)[name = string("transpose_156")];
            tensor<fp16, [1, 1024, 1024]> var_2224 = conv(dilations = var_2224_dilations_0, groups = var_2224_groups_0, pad = var_2224_pad_0, pad_type = var_2224_pad_type_0, strides = var_2224_strides_0, weight = squeeze_10, x = var_2216)[name = string("op_2224")];
            tensor<int32, [3]> var_2225 = const()[name = string("op_2225"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_65 = transpose(perm = var_2225, x = var_2224)[name = string("transpose_155")];
            tensor<fp16, [1, 1024, 1024]> x_367_cast_fp16 = add(x = hidden_states_21_cast_fp16, y = attn_out_65)[name = string("x_367_cast_fp16")];
            fp16 var_5_promoted_43_to_fp16 = const()[name = string("op_5_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_2231_cast_fp16 = pow(x = x_367_cast_fp16, y = var_5_promoted_43_to_fp16)[name = string("op_2231_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_87_cast_fp16_0 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_2231_cast_fp16)[name = string("var_87_cast_fp16")];
            fp16 var_2234_to_fp16 = const()[name = string("op_2234_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2235_cast_fp16 = add(x = var_87_cast_fp16_0, y = var_2234_to_fp16)[name = string("op_2235_cast_fp16")];
            fp32 var_2236_epsilon_0 = const()[name = string("op_2236_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2236_cast_fp16 = rsqrt(epsilon = var_2236_epsilon_0, x = var_2235_cast_fp16)[name = string("op_2236_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_371_cast_fp16 = mul(x = x_367_cast_fp16, y = var_2236_cast_fp16)[name = string("x_371_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_10_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_10_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1137516800)))];
            tensor<fp16, [1, 1024, 1024]> var_2239_cast_fp16 = mul(x = x_371_cast_fp16, y = encoder_layers_10_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_2239_cast_fp16")];
            tensor<int32, [3]> var_2244 = const()[name = string("op_2244"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_105_axes_0 = const()[name = string("input_105_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2245 = transpose(perm = var_2244, x = var_2239_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [1, 1024, 1, 1024]> input_105 = expand_dims(axes = input_105_axes_0, x = var_2245)[name = string("input_105")];
            string input_107_pad_type_0 = const()[name = string("input_107_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_107_strides_0 = const()[name = string("input_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_107_pad_0 = const()[name = string("input_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_107_dilations_0 = const()[name = string("input_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_107_groups_0 = const()[name = string("input_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_107 = conv(dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = encoder_layers_10_mlp_gate_proj_weight, x = input_105)[name = string("input_107")];
            string up_21_pad_type_0 = const()[name = string("up_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_21_strides_0 = const()[name = string("up_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_21_pad_0 = const()[name = string("up_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_21_dilations_0 = const()[name = string("up_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_21_groups_0 = const()[name = string("up_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_21 = conv(dilations = up_21_dilations_0, groups = up_21_groups_0, pad = up_21_pad_0, pad_type = up_21_pad_type_0, strides = up_21_strides_0, weight = encoder_layers_10_mlp_up_proj_weight, x = input_105)[name = string("up_21")];
            tensor<fp16, [1, 3072, 1, 1024]> var_2259 = silu(x = input_107)[name = string("op_2259")];
            tensor<fp16, [1, 3072, 1, 1024]> input_109 = mul(x = var_2259, y = up_21)[name = string("input_109")];
            string var_2266_pad_type_0 = const()[name = string("op_2266_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2266_strides_0 = const()[name = string("op_2266_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2266_pad_0 = const()[name = string("op_2266_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2266_dilations_0 = const()[name = string("op_2266_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2266_groups_0 = const()[name = string("op_2266_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2266 = conv(dilations = var_2266_dilations_0, groups = var_2266_groups_0, pad = var_2266_pad_0, pad_type = var_2266_pad_type_0, strides = var_2266_strides_0, weight = encoder_layers_10_mlp_down_proj_weight, x = input_109)[name = string("op_2266")];
            tensor<int32, [1]> var_2267_axes_0 = const()[name = string("op_2267_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2267 = squeeze(axes = var_2267_axes_0, x = var_2266)[name = string("op_2267")];
            tensor<int32, [3]> var_2268 = const()[name = string("op_2268"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_21 = transpose(perm = var_2268, x = var_2267)[name = string("transpose_153")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_23_cast_fp16 = add(x = x_367_cast_fp16, y = mlp_out_21)[name = string("hidden_states_23_cast_fp16")];
            fp16 var_5_promoted_44_to_fp16 = const()[name = string("op_5_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_2295_cast_fp16 = pow(x = hidden_states_23_cast_fp16, y = var_5_promoted_44_to_fp16)[name = string("op_2295_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_2295_cast_fp16)[name = string("var_89_cast_fp16")];
            fp16 var_2298_to_fp16 = const()[name = string("op_2298_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2299_cast_fp16 = add(x = var_89_cast_fp16, y = var_2298_to_fp16)[name = string("op_2299_cast_fp16")];
            fp32 var_2300_epsilon_0 = const()[name = string("op_2300_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2300_cast_fp16 = rsqrt(epsilon = var_2300_epsilon_0, x = var_2299_cast_fp16)[name = string("op_2300_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_377_cast_fp16 = mul(x = hidden_states_23_cast_fp16, y = var_2300_cast_fp16)[name = string("x_377_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_11_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_11_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1137518912)))];
            tensor<fp16, [1, 1024, 1024]> var_2303_cast_fp16 = mul(x = x_377_cast_fp16, y = encoder_layers_11_input_layernorm_weight_promoted_to_fp16)[name = string("op_2303_cast_fp16")];
            tensor<int32, [3]> var_2308 = const()[name = string("op_2308"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_111_axes_0 = const()[name = string("input_111_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2309 = transpose(perm = var_2308, x = var_2303_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 1024, 1, 1024]> input_111 = expand_dims(axes = input_111_axes_0, x = var_2309)[name = string("input_111")];
            string var_2316_pad_type_0 = const()[name = string("op_2316_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2316_strides_0 = const()[name = string("op_2316_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2316_pad_0 = const()[name = string("op_2316_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2316_dilations_0 = const()[name = string("op_2316_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2316_groups_0 = const()[name = string("op_2316_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_2316 = conv(dilations = var_2316_dilations_0, groups = var_2316_groups_0, pad = var_2316_pad_0, pad_type = var_2316_pad_type_0, strides = var_2316_strides_0, weight = encoder_layers_11_self_attn_q_proj_weight, x = input_111)[name = string("op_2316")];
            tensor<int32, [4]> var_2317 = const()[name = string("op_2317"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_2318 = reshape(shape = var_2317, x = var_2316)[name = string("op_2318")];
            tensor<int32, [4]> var_2319 = const()[name = string("op_2319"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2326_pad_type_0 = const()[name = string("op_2326_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2326_strides_0 = const()[name = string("op_2326_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2326_pad_0 = const()[name = string("op_2326_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2326_dilations_0 = const()[name = string("op_2326_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2326_groups_0 = const()[name = string("op_2326_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2326 = conv(dilations = var_2326_dilations_0, groups = var_2326_groups_0, pad = var_2326_pad_0, pad_type = var_2326_pad_type_0, strides = var_2326_strides_0, weight = encoder_layers_11_self_attn_k_proj_weight, x = input_111)[name = string("op_2326")];
            tensor<int32, [4]> var_2327 = const()[name = string("op_2327"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_2328 = reshape(shape = var_2327, x = var_2326)[name = string("op_2328")];
            tensor<int32, [4]> var_2329 = const()[name = string("op_2329"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2336_pad_type_0 = const()[name = string("op_2336_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2336_strides_0 = const()[name = string("op_2336_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2336_pad_0 = const()[name = string("op_2336_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2336_dilations_0 = const()[name = string("op_2336_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2336_groups_0 = const()[name = string("op_2336_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2336 = conv(dilations = var_2336_dilations_0, groups = var_2336_groups_0, pad = var_2336_pad_0, pad_type = var_2336_pad_type_0, strides = var_2336_strides_0, weight = encoder_layers_11_self_attn_v_proj_weight, x = input_111)[name = string("op_2336")];
            tensor<int32, [4]> var_2337 = const()[name = string("op_2337"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_2338 = reshape(shape = var_2337, x = var_2336)[name = string("op_2338")];
            tensor<int32, [4]> var_2339 = const()[name = string("op_2339"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_45_to_fp16 = const()[name = string("op_5_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_67 = transpose(perm = var_2319, x = var_2318)[name = string("transpose_151")];
            tensor<fp16, [1, 16, 1024, 128]> var_2345_cast_fp16 = pow(x = q_67, y = var_5_promoted_45_to_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_2345_cast_fp16)[name = string("var_91_cast_fp16")];
            fp16 var_2348_to_fp16 = const()[name = string("op_2348_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_2349_cast_fp16 = add(x = var_91_cast_fp16, y = var_2348_to_fp16)[name = string("op_2349_cast_fp16")];
            fp32 var_2350_epsilon_0 = const()[name = string("op_2350_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_2350_cast_fp16 = rsqrt(epsilon = var_2350_epsilon_0, x = var_2349_cast_fp16)[name = string("op_2350_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_385_cast_fp16 = mul(x = q_67, y = var_2350_cast_fp16)[name = string("x_385_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_69 = mul(x = x_385_cast_fp16, y = encoder_layers_11_self_attn_q_norm_weight)[name = string("q_69")];
            fp16 var_5_promoted_46_to_fp16 = const()[name = string("op_5_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_67 = transpose(perm = var_2329, x = var_2328)[name = string("transpose_150")];
            tensor<fp16, [1, 8, 1024, 128]> var_2358_cast_fp16 = pow(x = k_67, y = var_5_promoted_46_to_fp16)[name = string("op_2358_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_2358_cast_fp16)[name = string("var_93_cast_fp16")];
            fp16 var_2361_to_fp16 = const()[name = string("op_2361_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_2362_cast_fp16 = add(x = var_93_cast_fp16, y = var_2361_to_fp16)[name = string("op_2362_cast_fp16")];
            fp32 var_2363_epsilon_0 = const()[name = string("op_2363_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_2363_cast_fp16 = rsqrt(epsilon = var_2363_epsilon_0, x = var_2362_cast_fp16)[name = string("op_2363_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_391_cast_fp16 = mul(x = k_67, y = var_2363_cast_fp16)[name = string("x_391_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_69 = mul(x = x_391_cast_fp16, y = encoder_layers_11_self_attn_k_norm_weight)[name = string("k_69")];
            tensor<fp16, [1, 16, 1024, 128]> var_2367 = mul(x = q_69, y = cos)[name = string("op_2367")];
            tensor<int32, [2]> var_2368_split_sizes_0 = const()[name = string("op_2368_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2368_axis_0 = const()[name = string("op_2368_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_2368_0, tensor<fp16, [1, 16, 1024, 64]> var_2368_1 = split(axis = var_2368_axis_0, split_sizes = var_2368_split_sizes_0, x = q_69)[name = string("op_2368")];
            fp16 const_36_promoted = const()[name = string("const_36_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_2370 = mul(x = var_2368_1, y = const_36_promoted)[name = string("op_2370")];
            bool var_2372_interleave_0 = const()[name = string("op_2372_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_2372 = concat(axis = var_17, interleave = var_2372_interleave_0, values = (var_2370, var_2368_0))[name = string("op_2372")];
            tensor<fp16, [1, 16, 1024, 128]> var_2373 = mul(x = var_2372, y = sin)[name = string("op_2373")];
            tensor<fp16, [1, 16, 1024, 128]> query_23 = add(x = var_2367, y = var_2373)[name = string("query_23")];
            tensor<fp16, [1, 8, 1024, 128]> var_2375 = mul(x = k_69, y = cos)[name = string("op_2375")];
            tensor<int32, [2]> var_2376_split_sizes_0 = const()[name = string("op_2376_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2376_axis_0 = const()[name = string("op_2376_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_2376_0, tensor<fp16, [1, 8, 1024, 64]> var_2376_1 = split(axis = var_2376_axis_0, split_sizes = var_2376_split_sizes_0, x = k_69)[name = string("op_2376")];
            fp16 const_37_promoted = const()[name = string("const_37_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_2378 = mul(x = var_2376_1, y = const_37_promoted)[name = string("op_2378")];
            bool var_2380_interleave_0 = const()[name = string("op_2380_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_2380 = concat(axis = var_17, interleave = var_2380_interleave_0, values = (var_2378, var_2376_0))[name = string("op_2380")];
            tensor<fp16, [1, 8, 1024, 128]> var_2381 = mul(x = var_2380, y = sin)[name = string("op_2381")];
            tensor<fp16, [1, 8, 1024, 128]> x_393 = add(x = var_2375, y = var_2381)[name = string("x_393")];
            tensor<int32, [1]> var_2383_axes_0 = const()[name = string("op_2383_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2383 = expand_dims(axes = var_2383_axes_0, x = x_393)[name = string("op_2383")];
            tensor<int32, [5]> x_395_reps_0 = const()[name = string("x_395_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_395 = tile(reps = x_395_reps_0, x = var_2383)[name = string("x_395")];
            tensor<int32, [4]> var_2386 = const()[name = string("op_2386"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_23 = reshape(shape = var_2386, x = x_395)[name = string("key_23")];
            tensor<int32, [1]> var_2388_axes_0 = const()[name = string("op_2388_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_397 = transpose(perm = var_2339, x = var_2338)[name = string("transpose_149")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2388 = expand_dims(axes = var_2388_axes_0, x = x_397)[name = string("op_2388")];
            tensor<int32, [5]> x_399_reps_0 = const()[name = string("x_399_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_399 = tile(reps = x_399_reps_0, x = var_2388)[name = string("x_399")];
            tensor<int32, [4]> var_2391 = const()[name = string("op_2391"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_23 = reshape(shape = var_2391, x = x_399)[name = string("value_23")];
            bool var_2396_transpose_x_1 = const()[name = string("op_2396_transpose_x_1"), val = bool(false)];
            bool var_2396_transpose_y_1 = const()[name = string("op_2396_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_2396_cast_fp16 = matmul(transpose_x = var_2396_transpose_x_1, transpose_y = var_2396_transpose_y_1, x = query_23, y = key_23)[name = string("op_2396_cast_fp16")];
            fp16 var_2397_to_fp16 = const()[name = string("op_2397_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_67_cast_fp16 = mul(x = var_2396_cast_fp16, y = var_2397_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_2401_cast_fp16 = softmax(axis = var_17, x = attn_weights_69_cast_fp16)[name = string("op_2401_cast_fp16")];
            bool var_2405_transpose_x_0 = const()[name = string("op_2405_transpose_x_0"), val = bool(false)];
            bool var_2405_transpose_y_0 = const()[name = string("op_2405_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_2405_cast_fp16 = matmul(transpose_x = var_2405_transpose_x_0, transpose_y = var_2405_transpose_y_0, x = var_2401_cast_fp16, y = value_23)[name = string("op_2405_cast_fp16")];
            tensor<int32, [4]> var_2407 = const()[name = string("op_2407"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2410 = const()[name = string("op_2410"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_2408 = transpose(perm = var_2407, x = var_2405_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 1024, 2048]> attn_out_69 = reshape(shape = var_2410, x = var_2408)[name = string("attn_out_69")];
            tensor<int32, [3]> var_2412 = const()[name = string("op_2412"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_11 = const()[name = string("squeeze_11"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1137521024)))];
            string var_2421_pad_type_0 = const()[name = string("op_2421_pad_type_0"), val = string("valid")];
            int32 var_2421_groups_0 = const()[name = string("op_2421_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_2421_strides_0 = const()[name = string("op_2421_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_2421_pad_0 = const()[name = string("op_2421_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_2421_dilations_0 = const()[name = string("op_2421_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_2413 = transpose(perm = var_2412, x = attn_out_69)[name = string("transpose_147")];
            tensor<fp16, [1, 1024, 1024]> var_2421 = conv(dilations = var_2421_dilations_0, groups = var_2421_groups_0, pad = var_2421_pad_0, pad_type = var_2421_pad_type_0, strides = var_2421_strides_0, weight = squeeze_11, x = var_2413)[name = string("op_2421")];
            tensor<int32, [3]> var_2422 = const()[name = string("op_2422"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_71 = transpose(perm = var_2422, x = var_2421)[name = string("transpose_146")];
            tensor<fp16, [1, 1024, 1024]> x_401_cast_fp16 = add(x = hidden_states_23_cast_fp16, y = attn_out_71)[name = string("x_401_cast_fp16")];
            fp16 var_5_promoted_47_to_fp16 = const()[name = string("op_5_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_2428_cast_fp16 = pow(x = x_401_cast_fp16, y = var_5_promoted_47_to_fp16)[name = string("op_2428_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_2428_cast_fp16)[name = string("var_95_cast_fp16")];
            fp16 var_2431_to_fp16 = const()[name = string("op_2431_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2432_cast_fp16 = add(x = var_95_cast_fp16, y = var_2431_to_fp16)[name = string("op_2432_cast_fp16")];
            fp32 var_2433_epsilon_0 = const()[name = string("op_2433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2433_cast_fp16 = rsqrt(epsilon = var_2433_epsilon_0, x = var_2432_cast_fp16)[name = string("op_2433_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_405_cast_fp16 = mul(x = x_401_cast_fp16, y = var_2433_cast_fp16)[name = string("x_405_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_11_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_11_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1141715392)))];
            tensor<fp16, [1, 1024, 1024]> var_2436_cast_fp16 = mul(x = x_405_cast_fp16, y = encoder_layers_11_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_2436_cast_fp16")];
            tensor<int32, [3]> var_2441 = const()[name = string("op_2441"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_115_axes_0 = const()[name = string("input_115_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2442 = transpose(perm = var_2441, x = var_2436_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [1, 1024, 1, 1024]> input_115 = expand_dims(axes = input_115_axes_0, x = var_2442)[name = string("input_115")];
            string input_117_pad_type_0 = const()[name = string("input_117_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_117_strides_0 = const()[name = string("input_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_117_pad_0 = const()[name = string("input_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_117_dilations_0 = const()[name = string("input_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_117_groups_0 = const()[name = string("input_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_117 = conv(dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = encoder_layers_11_mlp_gate_proj_weight, x = input_115)[name = string("input_117")];
            string up_23_pad_type_0 = const()[name = string("up_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_23_strides_0 = const()[name = string("up_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_23_pad_0 = const()[name = string("up_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_23_dilations_0 = const()[name = string("up_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_23_groups_0 = const()[name = string("up_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_23 = conv(dilations = up_23_dilations_0, groups = up_23_groups_0, pad = up_23_pad_0, pad_type = up_23_pad_type_0, strides = up_23_strides_0, weight = encoder_layers_11_mlp_up_proj_weight, x = input_115)[name = string("up_23")];
            tensor<fp16, [1, 3072, 1, 1024]> var_2456 = silu(x = input_117)[name = string("op_2456")];
            tensor<fp16, [1, 3072, 1, 1024]> input_119 = mul(x = var_2456, y = up_23)[name = string("input_119")];
            string var_2463_pad_type_0 = const()[name = string("op_2463_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2463_strides_0 = const()[name = string("op_2463_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2463_pad_0 = const()[name = string("op_2463_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2463_dilations_0 = const()[name = string("op_2463_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2463_groups_0 = const()[name = string("op_2463_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2463 = conv(dilations = var_2463_dilations_0, groups = var_2463_groups_0, pad = var_2463_pad_0, pad_type = var_2463_pad_type_0, strides = var_2463_strides_0, weight = encoder_layers_11_mlp_down_proj_weight, x = input_119)[name = string("op_2463")];
            tensor<int32, [1]> var_2464_axes_0 = const()[name = string("op_2464_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2464 = squeeze(axes = var_2464_axes_0, x = var_2463)[name = string("op_2464")];
            tensor<int32, [3]> var_2465 = const()[name = string("op_2465"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_23 = transpose(perm = var_2465, x = var_2464)[name = string("transpose_144")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_25_cast_fp16 = add(x = x_401_cast_fp16, y = mlp_out_23)[name = string("hidden_states_25_cast_fp16")];
            fp16 var_5_promoted_48_to_fp16 = const()[name = string("op_5_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_2492_cast_fp16 = pow(x = hidden_states_25_cast_fp16, y = var_5_promoted_48_to_fp16)[name = string("op_2492_cast_fp16")];
            tensor<int32, [1]> var_97_axes_0 = const()[name = string("var_97_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_97_keep_dims_0 = const()[name = string("var_97_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_97_cast_fp16_0 = reduce_mean(axes = var_97_axes_0, keep_dims = var_97_keep_dims_0, x = var_2492_cast_fp16)[name = string("var_97_cast_fp16")];
            fp16 var_2495_to_fp16 = const()[name = string("op_2495_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2496_cast_fp16 = add(x = var_97_cast_fp16_0, y = var_2495_to_fp16)[name = string("op_2496_cast_fp16")];
            fp32 var_2497_epsilon_0 = const()[name = string("op_2497_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2497_cast_fp16 = rsqrt(epsilon = var_2497_epsilon_0, x = var_2496_cast_fp16)[name = string("op_2497_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_411_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = var_2497_cast_fp16)[name = string("x_411_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_12_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_12_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1141717504)))];
            tensor<fp16, [1, 1024, 1024]> var_2500_cast_fp16 = mul(x = x_411_cast_fp16, y = encoder_layers_12_input_layernorm_weight_promoted_to_fp16)[name = string("op_2500_cast_fp16")];
            tensor<int32, [3]> var_2505 = const()[name = string("op_2505"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_121_axes_0 = const()[name = string("input_121_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2506 = transpose(perm = var_2505, x = var_2500_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [1, 1024, 1, 1024]> input_121 = expand_dims(axes = input_121_axes_0, x = var_2506)[name = string("input_121")];
            string var_2513_pad_type_0 = const()[name = string("op_2513_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2513_strides_0 = const()[name = string("op_2513_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2513_pad_0 = const()[name = string("op_2513_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2513_dilations_0 = const()[name = string("op_2513_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2513_groups_0 = const()[name = string("op_2513_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_2513 = conv(dilations = var_2513_dilations_0, groups = var_2513_groups_0, pad = var_2513_pad_0, pad_type = var_2513_pad_type_0, strides = var_2513_strides_0, weight = encoder_layers_12_self_attn_q_proj_weight, x = input_121)[name = string("op_2513")];
            tensor<int32, [4]> var_2514 = const()[name = string("op_2514"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_2515 = reshape(shape = var_2514, x = var_2513)[name = string("op_2515")];
            tensor<int32, [4]> var_2516 = const()[name = string("op_2516"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2523_pad_type_0 = const()[name = string("op_2523_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2523_strides_0 = const()[name = string("op_2523_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2523_pad_0 = const()[name = string("op_2523_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2523_dilations_0 = const()[name = string("op_2523_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2523_groups_0 = const()[name = string("op_2523_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2523 = conv(dilations = var_2523_dilations_0, groups = var_2523_groups_0, pad = var_2523_pad_0, pad_type = var_2523_pad_type_0, strides = var_2523_strides_0, weight = encoder_layers_12_self_attn_k_proj_weight, x = input_121)[name = string("op_2523")];
            tensor<int32, [4]> var_2524 = const()[name = string("op_2524"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_2525 = reshape(shape = var_2524, x = var_2523)[name = string("op_2525")];
            tensor<int32, [4]> var_2526 = const()[name = string("op_2526"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2533_pad_type_0 = const()[name = string("op_2533_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2533_strides_0 = const()[name = string("op_2533_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2533_pad_0 = const()[name = string("op_2533_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2533_dilations_0 = const()[name = string("op_2533_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2533_groups_0 = const()[name = string("op_2533_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2533 = conv(dilations = var_2533_dilations_0, groups = var_2533_groups_0, pad = var_2533_pad_0, pad_type = var_2533_pad_type_0, strides = var_2533_strides_0, weight = encoder_layers_12_self_attn_v_proj_weight, x = input_121)[name = string("op_2533")];
            tensor<int32, [4]> var_2534 = const()[name = string("op_2534"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_2535 = reshape(shape = var_2534, x = var_2533)[name = string("op_2535")];
            tensor<int32, [4]> var_2536 = const()[name = string("op_2536"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_49_to_fp16 = const()[name = string("op_5_promoted_49_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_73 = transpose(perm = var_2516, x = var_2515)[name = string("transpose_142")];
            tensor<fp16, [1, 16, 1024, 128]> var_2542_cast_fp16 = pow(x = q_73, y = var_5_promoted_49_to_fp16)[name = string("op_2542_cast_fp16")];
            tensor<int32, [1]> var_99_axes_0 = const()[name = string("var_99_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_99_keep_dims_0 = const()[name = string("var_99_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_99_cast_fp16 = reduce_mean(axes = var_99_axes_0, keep_dims = var_99_keep_dims_0, x = var_2542_cast_fp16)[name = string("var_99_cast_fp16")];
            fp16 var_2545_to_fp16 = const()[name = string("op_2545_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_2546_cast_fp16 = add(x = var_99_cast_fp16, y = var_2545_to_fp16)[name = string("op_2546_cast_fp16")];
            fp32 var_2547_epsilon_0 = const()[name = string("op_2547_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_2547_cast_fp16 = rsqrt(epsilon = var_2547_epsilon_0, x = var_2546_cast_fp16)[name = string("op_2547_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_419_cast_fp16 = mul(x = q_73, y = var_2547_cast_fp16)[name = string("x_419_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_75 = mul(x = x_419_cast_fp16, y = encoder_layers_12_self_attn_q_norm_weight)[name = string("q_75")];
            fp16 var_5_promoted_50_to_fp16 = const()[name = string("op_5_promoted_50_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_73 = transpose(perm = var_2526, x = var_2525)[name = string("transpose_141")];
            tensor<fp16, [1, 8, 1024, 128]> var_2555_cast_fp16 = pow(x = k_73, y = var_5_promoted_50_to_fp16)[name = string("op_2555_cast_fp16")];
            tensor<int32, [1]> var_101_axes_0 = const()[name = string("var_101_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_101_keep_dims_0 = const()[name = string("var_101_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_101_cast_fp16 = reduce_mean(axes = var_101_axes_0, keep_dims = var_101_keep_dims_0, x = var_2555_cast_fp16)[name = string("var_101_cast_fp16")];
            fp16 var_2558_to_fp16 = const()[name = string("op_2558_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_2559_cast_fp16 = add(x = var_101_cast_fp16, y = var_2558_to_fp16)[name = string("op_2559_cast_fp16")];
            fp32 var_2560_epsilon_0 = const()[name = string("op_2560_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_2560_cast_fp16 = rsqrt(epsilon = var_2560_epsilon_0, x = var_2559_cast_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_425_cast_fp16 = mul(x = k_73, y = var_2560_cast_fp16)[name = string("x_425_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_75 = mul(x = x_425_cast_fp16, y = encoder_layers_12_self_attn_k_norm_weight)[name = string("k_75")];
            tensor<fp16, [1, 16, 1024, 128]> var_2564 = mul(x = q_75, y = cos)[name = string("op_2564")];
            tensor<int32, [2]> var_2565_split_sizes_0 = const()[name = string("op_2565_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2565_axis_0 = const()[name = string("op_2565_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_2565_0, tensor<fp16, [1, 16, 1024, 64]> var_2565_1 = split(axis = var_2565_axis_0, split_sizes = var_2565_split_sizes_0, x = q_75)[name = string("op_2565")];
            fp16 const_39_promoted = const()[name = string("const_39_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_2567 = mul(x = var_2565_1, y = const_39_promoted)[name = string("op_2567")];
            bool var_2569_interleave_0 = const()[name = string("op_2569_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_2569 = concat(axis = var_17, interleave = var_2569_interleave_0, values = (var_2567, var_2565_0))[name = string("op_2569")];
            tensor<fp16, [1, 16, 1024, 128]> var_2570 = mul(x = var_2569, y = sin)[name = string("op_2570")];
            tensor<fp16, [1, 16, 1024, 128]> query_25 = add(x = var_2564, y = var_2570)[name = string("query_25")];
            tensor<fp16, [1, 8, 1024, 128]> var_2572 = mul(x = k_75, y = cos)[name = string("op_2572")];
            tensor<int32, [2]> var_2573_split_sizes_0 = const()[name = string("op_2573_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2573_axis_0 = const()[name = string("op_2573_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_2573_0, tensor<fp16, [1, 8, 1024, 64]> var_2573_1 = split(axis = var_2573_axis_0, split_sizes = var_2573_split_sizes_0, x = k_75)[name = string("op_2573")];
            fp16 const_40_promoted = const()[name = string("const_40_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_2575 = mul(x = var_2573_1, y = const_40_promoted)[name = string("op_2575")];
            bool var_2577_interleave_0 = const()[name = string("op_2577_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_2577 = concat(axis = var_17, interleave = var_2577_interleave_0, values = (var_2575, var_2573_0))[name = string("op_2577")];
            tensor<fp16, [1, 8, 1024, 128]> var_2578 = mul(x = var_2577, y = sin)[name = string("op_2578")];
            tensor<fp16, [1, 8, 1024, 128]> x_427 = add(x = var_2572, y = var_2578)[name = string("x_427")];
            tensor<int32, [1]> var_2580_axes_0 = const()[name = string("op_2580_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2580 = expand_dims(axes = var_2580_axes_0, x = x_427)[name = string("op_2580")];
            tensor<int32, [5]> x_429_reps_0 = const()[name = string("x_429_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_429 = tile(reps = x_429_reps_0, x = var_2580)[name = string("x_429")];
            tensor<int32, [4]> var_2583 = const()[name = string("op_2583"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_25 = reshape(shape = var_2583, x = x_429)[name = string("key_25")];
            tensor<int32, [1]> var_2585_axes_0 = const()[name = string("op_2585_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_431 = transpose(perm = var_2536, x = var_2535)[name = string("transpose_140")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2585 = expand_dims(axes = var_2585_axes_0, x = x_431)[name = string("op_2585")];
            tensor<int32, [5]> x_433_reps_0 = const()[name = string("x_433_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_433 = tile(reps = x_433_reps_0, x = var_2585)[name = string("x_433")];
            tensor<int32, [4]> var_2588 = const()[name = string("op_2588"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_25 = reshape(shape = var_2588, x = x_433)[name = string("value_25")];
            bool var_2593_transpose_x_1 = const()[name = string("op_2593_transpose_x_1"), val = bool(false)];
            bool var_2593_transpose_y_1 = const()[name = string("op_2593_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_2593_cast_fp16 = matmul(transpose_x = var_2593_transpose_x_1, transpose_y = var_2593_transpose_y_1, x = query_25, y = key_25)[name = string("op_2593_cast_fp16")];
            fp16 var_2594_to_fp16 = const()[name = string("op_2594_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_73_cast_fp16 = mul(x = var_2593_cast_fp16, y = var_2594_to_fp16)[name = string("attn_weights_73_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_75_cast_fp16 = add(x = attn_weights_73_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_2598_cast_fp16 = softmax(axis = var_17, x = attn_weights_75_cast_fp16)[name = string("op_2598_cast_fp16")];
            bool var_2602_transpose_x_0 = const()[name = string("op_2602_transpose_x_0"), val = bool(false)];
            bool var_2602_transpose_y_0 = const()[name = string("op_2602_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_2602_cast_fp16 = matmul(transpose_x = var_2602_transpose_x_0, transpose_y = var_2602_transpose_y_0, x = var_2598_cast_fp16, y = value_25)[name = string("op_2602_cast_fp16")];
            tensor<int32, [4]> var_2604 = const()[name = string("op_2604"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2607 = const()[name = string("op_2607"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_2605 = transpose(perm = var_2604, x = var_2602_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 1024, 2048]> attn_out_75 = reshape(shape = var_2607, x = var_2605)[name = string("attn_out_75")];
            tensor<int32, [3]> var_2609 = const()[name = string("op_2609"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_12 = const()[name = string("squeeze_12"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1141719616)))];
            string var_2618_pad_type_0 = const()[name = string("op_2618_pad_type_0"), val = string("valid")];
            int32 var_2618_groups_0 = const()[name = string("op_2618_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_2618_strides_0 = const()[name = string("op_2618_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_2618_pad_0 = const()[name = string("op_2618_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_2618_dilations_0 = const()[name = string("op_2618_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_2610 = transpose(perm = var_2609, x = attn_out_75)[name = string("transpose_138")];
            tensor<fp16, [1, 1024, 1024]> var_2618 = conv(dilations = var_2618_dilations_0, groups = var_2618_groups_0, pad = var_2618_pad_0, pad_type = var_2618_pad_type_0, strides = var_2618_strides_0, weight = squeeze_12, x = var_2610)[name = string("op_2618")];
            tensor<int32, [3]> var_2619 = const()[name = string("op_2619"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_77 = transpose(perm = var_2619, x = var_2618)[name = string("transpose_137")];
            tensor<fp16, [1, 1024, 1024]> x_435_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = attn_out_77)[name = string("x_435_cast_fp16")];
            fp16 var_5_promoted_51_to_fp16 = const()[name = string("op_5_promoted_51_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_2625_cast_fp16 = pow(x = x_435_cast_fp16, y = var_5_promoted_51_to_fp16)[name = string("op_2625_cast_fp16")];
            tensor<int32, [1]> var_103_axes_0 = const()[name = string("var_103_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_103_keep_dims_0 = const()[name = string("var_103_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_103_cast_fp16 = reduce_mean(axes = var_103_axes_0, keep_dims = var_103_keep_dims_0, x = var_2625_cast_fp16)[name = string("var_103_cast_fp16")];
            fp16 var_2628_to_fp16 = const()[name = string("op_2628_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2629_cast_fp16 = add(x = var_103_cast_fp16, y = var_2628_to_fp16)[name = string("op_2629_cast_fp16")];
            fp32 var_2630_epsilon_0 = const()[name = string("op_2630_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2630_cast_fp16 = rsqrt(epsilon = var_2630_epsilon_0, x = var_2629_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_439_cast_fp16 = mul(x = x_435_cast_fp16, y = var_2630_cast_fp16)[name = string("x_439_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_12_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_12_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1145913984)))];
            tensor<fp16, [1, 1024, 1024]> var_2633_cast_fp16 = mul(x = x_439_cast_fp16, y = encoder_layers_12_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_2633_cast_fp16")];
            tensor<int32, [3]> var_2638 = const()[name = string("op_2638"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_125_axes_0 = const()[name = string("input_125_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2639 = transpose(perm = var_2638, x = var_2633_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 1024, 1, 1024]> input_125 = expand_dims(axes = input_125_axes_0, x = var_2639)[name = string("input_125")];
            string input_127_pad_type_0 = const()[name = string("input_127_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_127_strides_0 = const()[name = string("input_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_127_pad_0 = const()[name = string("input_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_127_dilations_0 = const()[name = string("input_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_127_groups_0 = const()[name = string("input_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_127 = conv(dilations = input_127_dilations_0, groups = input_127_groups_0, pad = input_127_pad_0, pad_type = input_127_pad_type_0, strides = input_127_strides_0, weight = encoder_layers_12_mlp_gate_proj_weight, x = input_125)[name = string("input_127")];
            string up_25_pad_type_0 = const()[name = string("up_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_25_strides_0 = const()[name = string("up_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_25_pad_0 = const()[name = string("up_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_25_dilations_0 = const()[name = string("up_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_25_groups_0 = const()[name = string("up_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_25 = conv(dilations = up_25_dilations_0, groups = up_25_groups_0, pad = up_25_pad_0, pad_type = up_25_pad_type_0, strides = up_25_strides_0, weight = encoder_layers_12_mlp_up_proj_weight, x = input_125)[name = string("up_25")];
            tensor<fp16, [1, 3072, 1, 1024]> var_2653 = silu(x = input_127)[name = string("op_2653")];
            tensor<fp16, [1, 3072, 1, 1024]> input_129 = mul(x = var_2653, y = up_25)[name = string("input_129")];
            string var_2660_pad_type_0 = const()[name = string("op_2660_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2660_strides_0 = const()[name = string("op_2660_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2660_pad_0 = const()[name = string("op_2660_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2660_dilations_0 = const()[name = string("op_2660_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2660_groups_0 = const()[name = string("op_2660_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2660 = conv(dilations = var_2660_dilations_0, groups = var_2660_groups_0, pad = var_2660_pad_0, pad_type = var_2660_pad_type_0, strides = var_2660_strides_0, weight = encoder_layers_12_mlp_down_proj_weight, x = input_129)[name = string("op_2660")];
            tensor<int32, [1]> var_2661_axes_0 = const()[name = string("op_2661_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2661 = squeeze(axes = var_2661_axes_0, x = var_2660)[name = string("op_2661")];
            tensor<int32, [3]> var_2662 = const()[name = string("op_2662"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_25 = transpose(perm = var_2662, x = var_2661)[name = string("transpose_135")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_27_cast_fp16 = add(x = x_435_cast_fp16, y = mlp_out_25)[name = string("hidden_states_27_cast_fp16")];
            fp16 var_5_promoted_52_to_fp16 = const()[name = string("op_5_promoted_52_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_2689_cast_fp16 = pow(x = hidden_states_27_cast_fp16, y = var_5_promoted_52_to_fp16)[name = string("op_2689_cast_fp16")];
            tensor<int32, [1]> var_105_axes_0 = const()[name = string("var_105_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_105_keep_dims_0 = const()[name = string("var_105_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_105_cast_fp16 = reduce_mean(axes = var_105_axes_0, keep_dims = var_105_keep_dims_0, x = var_2689_cast_fp16)[name = string("var_105_cast_fp16")];
            fp16 var_2692_to_fp16 = const()[name = string("op_2692_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2693_cast_fp16 = add(x = var_105_cast_fp16, y = var_2692_to_fp16)[name = string("op_2693_cast_fp16")];
            fp32 var_2694_epsilon_0 = const()[name = string("op_2694_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2694_cast_fp16 = rsqrt(epsilon = var_2694_epsilon_0, x = var_2693_cast_fp16)[name = string("op_2694_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_445_cast_fp16 = mul(x = hidden_states_27_cast_fp16, y = var_2694_cast_fp16)[name = string("x_445_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_13_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_13_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1145916096)))];
            tensor<fp16, [1, 1024, 1024]> var_2697_cast_fp16 = mul(x = x_445_cast_fp16, y = encoder_layers_13_input_layernorm_weight_promoted_to_fp16)[name = string("op_2697_cast_fp16")];
            tensor<int32, [3]> var_2702 = const()[name = string("op_2702"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2703 = transpose(perm = var_2702, x = var_2697_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 1024, 1, 1024]> input_131 = expand_dims(axes = input_131_axes_0, x = var_2703)[name = string("input_131")];
            string var_2710_pad_type_0 = const()[name = string("op_2710_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2710_strides_0 = const()[name = string("op_2710_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2710_pad_0 = const()[name = string("op_2710_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2710_dilations_0 = const()[name = string("op_2710_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2710_groups_0 = const()[name = string("op_2710_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_2710 = conv(dilations = var_2710_dilations_0, groups = var_2710_groups_0, pad = var_2710_pad_0, pad_type = var_2710_pad_type_0, strides = var_2710_strides_0, weight = encoder_layers_13_self_attn_q_proj_weight, x = input_131)[name = string("op_2710")];
            tensor<int32, [4]> var_2711 = const()[name = string("op_2711"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_2712 = reshape(shape = var_2711, x = var_2710)[name = string("op_2712")];
            tensor<int32, [4]> var_2713 = const()[name = string("op_2713"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2720_pad_type_0 = const()[name = string("op_2720_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2720_strides_0 = const()[name = string("op_2720_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2720_pad_0 = const()[name = string("op_2720_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2720_dilations_0 = const()[name = string("op_2720_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2720_groups_0 = const()[name = string("op_2720_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2720 = conv(dilations = var_2720_dilations_0, groups = var_2720_groups_0, pad = var_2720_pad_0, pad_type = var_2720_pad_type_0, strides = var_2720_strides_0, weight = encoder_layers_13_self_attn_k_proj_weight, x = input_131)[name = string("op_2720")];
            tensor<int32, [4]> var_2721 = const()[name = string("op_2721"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_2722 = reshape(shape = var_2721, x = var_2720)[name = string("op_2722")];
            tensor<int32, [4]> var_2723 = const()[name = string("op_2723"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2730_pad_type_0 = const()[name = string("op_2730_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2730_strides_0 = const()[name = string("op_2730_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2730_pad_0 = const()[name = string("op_2730_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2730_dilations_0 = const()[name = string("op_2730_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2730_groups_0 = const()[name = string("op_2730_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2730 = conv(dilations = var_2730_dilations_0, groups = var_2730_groups_0, pad = var_2730_pad_0, pad_type = var_2730_pad_type_0, strides = var_2730_strides_0, weight = encoder_layers_13_self_attn_v_proj_weight, x = input_131)[name = string("op_2730")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_2732 = reshape(shape = var_2731, x = var_2730)[name = string("op_2732")];
            tensor<int32, [4]> var_2733 = const()[name = string("op_2733"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_53_to_fp16 = const()[name = string("op_5_promoted_53_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_79 = transpose(perm = var_2713, x = var_2712)[name = string("transpose_133")];
            tensor<fp16, [1, 16, 1024, 128]> var_2739_cast_fp16 = pow(x = q_79, y = var_5_promoted_53_to_fp16)[name = string("op_2739_cast_fp16")];
            tensor<int32, [1]> var_107_axes_0 = const()[name = string("var_107_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_107_keep_dims_0 = const()[name = string("var_107_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_107_cast_fp16 = reduce_mean(axes = var_107_axes_0, keep_dims = var_107_keep_dims_0, x = var_2739_cast_fp16)[name = string("var_107_cast_fp16")];
            fp16 var_2742_to_fp16 = const()[name = string("op_2742_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_2743_cast_fp16 = add(x = var_107_cast_fp16, y = var_2742_to_fp16)[name = string("op_2743_cast_fp16")];
            fp32 var_2744_epsilon_0 = const()[name = string("op_2744_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_2744_cast_fp16 = rsqrt(epsilon = var_2744_epsilon_0, x = var_2743_cast_fp16)[name = string("op_2744_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_453_cast_fp16 = mul(x = q_79, y = var_2744_cast_fp16)[name = string("x_453_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_81 = mul(x = x_453_cast_fp16, y = encoder_layers_13_self_attn_q_norm_weight)[name = string("q_81")];
            fp16 var_5_promoted_54_to_fp16 = const()[name = string("op_5_promoted_54_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_79 = transpose(perm = var_2723, x = var_2722)[name = string("transpose_132")];
            tensor<fp16, [1, 8, 1024, 128]> var_2752_cast_fp16 = pow(x = k_79, y = var_5_promoted_54_to_fp16)[name = string("op_2752_cast_fp16")];
            tensor<int32, [1]> var_109_axes_0 = const()[name = string("var_109_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_109_keep_dims_0 = const()[name = string("var_109_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_109_cast_fp16 = reduce_mean(axes = var_109_axes_0, keep_dims = var_109_keep_dims_0, x = var_2752_cast_fp16)[name = string("var_109_cast_fp16")];
            fp16 var_2755_to_fp16 = const()[name = string("op_2755_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_2756_cast_fp16 = add(x = var_109_cast_fp16, y = var_2755_to_fp16)[name = string("op_2756_cast_fp16")];
            fp32 var_2757_epsilon_0 = const()[name = string("op_2757_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_2757_cast_fp16 = rsqrt(epsilon = var_2757_epsilon_0, x = var_2756_cast_fp16)[name = string("op_2757_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_459_cast_fp16 = mul(x = k_79, y = var_2757_cast_fp16)[name = string("x_459_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_81 = mul(x = x_459_cast_fp16, y = encoder_layers_13_self_attn_k_norm_weight)[name = string("k_81")];
            tensor<fp16, [1, 16, 1024, 128]> var_2761 = mul(x = q_81, y = cos)[name = string("op_2761")];
            tensor<int32, [2]> var_2762_split_sizes_0 = const()[name = string("op_2762_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2762_axis_0 = const()[name = string("op_2762_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_2762_0, tensor<fp16, [1, 16, 1024, 64]> var_2762_1 = split(axis = var_2762_axis_0, split_sizes = var_2762_split_sizes_0, x = q_81)[name = string("op_2762")];
            fp16 const_42_promoted = const()[name = string("const_42_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_2764 = mul(x = var_2762_1, y = const_42_promoted)[name = string("op_2764")];
            bool var_2766_interleave_0 = const()[name = string("op_2766_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_2766 = concat(axis = var_17, interleave = var_2766_interleave_0, values = (var_2764, var_2762_0))[name = string("op_2766")];
            tensor<fp16, [1, 16, 1024, 128]> var_2767 = mul(x = var_2766, y = sin)[name = string("op_2767")];
            tensor<fp16, [1, 16, 1024, 128]> query_27 = add(x = var_2761, y = var_2767)[name = string("query_27")];
            tensor<fp16, [1, 8, 1024, 128]> var_2769 = mul(x = k_81, y = cos)[name = string("op_2769")];
            tensor<int32, [2]> var_2770_split_sizes_0 = const()[name = string("op_2770_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2770_axis_0 = const()[name = string("op_2770_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_2770_0, tensor<fp16, [1, 8, 1024, 64]> var_2770_1 = split(axis = var_2770_axis_0, split_sizes = var_2770_split_sizes_0, x = k_81)[name = string("op_2770")];
            fp16 const_43_promoted = const()[name = string("const_43_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_2772 = mul(x = var_2770_1, y = const_43_promoted)[name = string("op_2772")];
            bool var_2774_interleave_0 = const()[name = string("op_2774_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_2774 = concat(axis = var_17, interleave = var_2774_interleave_0, values = (var_2772, var_2770_0))[name = string("op_2774")];
            tensor<fp16, [1, 8, 1024, 128]> var_2775 = mul(x = var_2774, y = sin)[name = string("op_2775")];
            tensor<fp16, [1, 8, 1024, 128]> x_461 = add(x = var_2769, y = var_2775)[name = string("x_461")];
            tensor<int32, [1]> var_2777_axes_0 = const()[name = string("op_2777_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2777 = expand_dims(axes = var_2777_axes_0, x = x_461)[name = string("op_2777")];
            tensor<int32, [5]> x_463_reps_0 = const()[name = string("x_463_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_463 = tile(reps = x_463_reps_0, x = var_2777)[name = string("x_463")];
            tensor<int32, [4]> var_2780 = const()[name = string("op_2780"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_27 = reshape(shape = var_2780, x = x_463)[name = string("key_27")];
            tensor<int32, [1]> var_2782_axes_0 = const()[name = string("op_2782_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_465 = transpose(perm = var_2733, x = var_2732)[name = string("transpose_131")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2782 = expand_dims(axes = var_2782_axes_0, x = x_465)[name = string("op_2782")];
            tensor<int32, [5]> x_467_reps_0 = const()[name = string("x_467_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_467 = tile(reps = x_467_reps_0, x = var_2782)[name = string("x_467")];
            tensor<int32, [4]> var_2785 = const()[name = string("op_2785"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_27 = reshape(shape = var_2785, x = x_467)[name = string("value_27")];
            bool var_2790_transpose_x_1 = const()[name = string("op_2790_transpose_x_1"), val = bool(false)];
            bool var_2790_transpose_y_1 = const()[name = string("op_2790_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_2790_cast_fp16 = matmul(transpose_x = var_2790_transpose_x_1, transpose_y = var_2790_transpose_y_1, x = query_27, y = key_27)[name = string("op_2790_cast_fp16")];
            fp16 var_2791_to_fp16 = const()[name = string("op_2791_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_79_cast_fp16 = mul(x = var_2790_cast_fp16, y = var_2791_to_fp16)[name = string("attn_weights_79_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_81_cast_fp16 = add(x = attn_weights_79_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_81_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_2795_cast_fp16 = softmax(axis = var_17, x = attn_weights_81_cast_fp16)[name = string("op_2795_cast_fp16")];
            bool var_2799_transpose_x_0 = const()[name = string("op_2799_transpose_x_0"), val = bool(false)];
            bool var_2799_transpose_y_0 = const()[name = string("op_2799_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_2799_cast_fp16 = matmul(transpose_x = var_2799_transpose_x_0, transpose_y = var_2799_transpose_y_0, x = var_2795_cast_fp16, y = value_27)[name = string("op_2799_cast_fp16")];
            tensor<int32, [4]> var_2801 = const()[name = string("op_2801"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2804 = const()[name = string("op_2804"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_2802 = transpose(perm = var_2801, x = var_2799_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 1024, 2048]> attn_out_81 = reshape(shape = var_2804, x = var_2802)[name = string("attn_out_81")];
            tensor<int32, [3]> var_2806 = const()[name = string("op_2806"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_13 = const()[name = string("squeeze_13"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1145918208)))];
            string var_2815_pad_type_0 = const()[name = string("op_2815_pad_type_0"), val = string("valid")];
            int32 var_2815_groups_0 = const()[name = string("op_2815_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_2815_strides_0 = const()[name = string("op_2815_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_2815_pad_0 = const()[name = string("op_2815_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_2815_dilations_0 = const()[name = string("op_2815_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_2807 = transpose(perm = var_2806, x = attn_out_81)[name = string("transpose_129")];
            tensor<fp16, [1, 1024, 1024]> var_2815 = conv(dilations = var_2815_dilations_0, groups = var_2815_groups_0, pad = var_2815_pad_0, pad_type = var_2815_pad_type_0, strides = var_2815_strides_0, weight = squeeze_13, x = var_2807)[name = string("op_2815")];
            tensor<int32, [3]> var_2816 = const()[name = string("op_2816"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_83 = transpose(perm = var_2816, x = var_2815)[name = string("transpose_128")];
            tensor<fp16, [1, 1024, 1024]> x_469_cast_fp16 = add(x = hidden_states_27_cast_fp16, y = attn_out_83)[name = string("x_469_cast_fp16")];
            fp16 var_5_promoted_55_to_fp16 = const()[name = string("op_5_promoted_55_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_2822_cast_fp16 = pow(x = x_469_cast_fp16, y = var_5_promoted_55_to_fp16)[name = string("op_2822_cast_fp16")];
            tensor<int32, [1]> var_111_axes_0 = const()[name = string("var_111_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_111_keep_dims_0 = const()[name = string("var_111_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_111_cast_fp16 = reduce_mean(axes = var_111_axes_0, keep_dims = var_111_keep_dims_0, x = var_2822_cast_fp16)[name = string("var_111_cast_fp16")];
            fp16 var_2825_to_fp16 = const()[name = string("op_2825_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2826_cast_fp16 = add(x = var_111_cast_fp16, y = var_2825_to_fp16)[name = string("op_2826_cast_fp16")];
            fp32 var_2827_epsilon_0 = const()[name = string("op_2827_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2827_cast_fp16 = rsqrt(epsilon = var_2827_epsilon_0, x = var_2826_cast_fp16)[name = string("op_2827_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_473_cast_fp16 = mul(x = x_469_cast_fp16, y = var_2827_cast_fp16)[name = string("x_473_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_13_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_13_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1150112576)))];
            tensor<fp16, [1, 1024, 1024]> var_2830_cast_fp16 = mul(x = x_473_cast_fp16, y = encoder_layers_13_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_2830_cast_fp16")];
            tensor<int32, [3]> var_2835 = const()[name = string("op_2835"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_135_axes_0 = const()[name = string("input_135_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2836 = transpose(perm = var_2835, x = var_2830_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 1024, 1, 1024]> input_135 = expand_dims(axes = input_135_axes_0, x = var_2836)[name = string("input_135")];
            string input_137_pad_type_0 = const()[name = string("input_137_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_137_strides_0 = const()[name = string("input_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_137_pad_0 = const()[name = string("input_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_137_dilations_0 = const()[name = string("input_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_137_groups_0 = const()[name = string("input_137_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_137 = conv(dilations = input_137_dilations_0, groups = input_137_groups_0, pad = input_137_pad_0, pad_type = input_137_pad_type_0, strides = input_137_strides_0, weight = encoder_layers_13_mlp_gate_proj_weight, x = input_135)[name = string("input_137")];
            string up_27_pad_type_0 = const()[name = string("up_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_27_strides_0 = const()[name = string("up_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_27_pad_0 = const()[name = string("up_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_27_dilations_0 = const()[name = string("up_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_27_groups_0 = const()[name = string("up_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_27 = conv(dilations = up_27_dilations_0, groups = up_27_groups_0, pad = up_27_pad_0, pad_type = up_27_pad_type_0, strides = up_27_strides_0, weight = encoder_layers_13_mlp_up_proj_weight, x = input_135)[name = string("up_27")];
            tensor<fp16, [1, 3072, 1, 1024]> var_2850 = silu(x = input_137)[name = string("op_2850")];
            tensor<fp16, [1, 3072, 1, 1024]> input_139 = mul(x = var_2850, y = up_27)[name = string("input_139")];
            string var_2857_pad_type_0 = const()[name = string("op_2857_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2857_strides_0 = const()[name = string("op_2857_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2857_pad_0 = const()[name = string("op_2857_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2857_dilations_0 = const()[name = string("op_2857_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2857_groups_0 = const()[name = string("op_2857_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2857 = conv(dilations = var_2857_dilations_0, groups = var_2857_groups_0, pad = var_2857_pad_0, pad_type = var_2857_pad_type_0, strides = var_2857_strides_0, weight = encoder_layers_13_mlp_down_proj_weight, x = input_139)[name = string("op_2857")];
            tensor<int32, [1]> var_2858_axes_0 = const()[name = string("op_2858_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2858 = squeeze(axes = var_2858_axes_0, x = var_2857)[name = string("op_2858")];
            tensor<int32, [3]> var_2859 = const()[name = string("op_2859"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_27 = transpose(perm = var_2859, x = var_2858)[name = string("transpose_126")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_29_cast_fp16 = add(x = x_469_cast_fp16, y = mlp_out_27)[name = string("hidden_states_29_cast_fp16")];
            fp16 var_5_promoted_56_to_fp16 = const()[name = string("op_5_promoted_56_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_2886_cast_fp16 = pow(x = hidden_states_29_cast_fp16, y = var_5_promoted_56_to_fp16)[name = string("op_2886_cast_fp16")];
            tensor<int32, [1]> var_113_axes_0 = const()[name = string("var_113_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_113_keep_dims_0 = const()[name = string("var_113_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_113_cast_fp16 = reduce_mean(axes = var_113_axes_0, keep_dims = var_113_keep_dims_0, x = var_2886_cast_fp16)[name = string("var_113_cast_fp16")];
            fp16 var_2889_to_fp16 = const()[name = string("op_2889_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2890_cast_fp16 = add(x = var_113_cast_fp16, y = var_2889_to_fp16)[name = string("op_2890_cast_fp16")];
            fp32 var_2891_epsilon_0 = const()[name = string("op_2891_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2891_cast_fp16 = rsqrt(epsilon = var_2891_epsilon_0, x = var_2890_cast_fp16)[name = string("op_2891_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_479_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = var_2891_cast_fp16)[name = string("x_479_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_14_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_14_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1150114688)))];
            tensor<fp16, [1, 1024, 1024]> var_2894_cast_fp16 = mul(x = x_479_cast_fp16, y = encoder_layers_14_input_layernorm_weight_promoted_to_fp16)[name = string("op_2894_cast_fp16")];
            tensor<int32, [3]> var_2899 = const()[name = string("op_2899"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_141_axes_0 = const()[name = string("input_141_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2900 = transpose(perm = var_2899, x = var_2894_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 1024, 1, 1024]> input_141 = expand_dims(axes = input_141_axes_0, x = var_2900)[name = string("input_141")];
            string var_2907_pad_type_0 = const()[name = string("op_2907_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2907_strides_0 = const()[name = string("op_2907_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2907_pad_0 = const()[name = string("op_2907_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2907_dilations_0 = const()[name = string("op_2907_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2907_groups_0 = const()[name = string("op_2907_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_2907 = conv(dilations = var_2907_dilations_0, groups = var_2907_groups_0, pad = var_2907_pad_0, pad_type = var_2907_pad_type_0, strides = var_2907_strides_0, weight = encoder_layers_14_self_attn_q_proj_weight, x = input_141)[name = string("op_2907")];
            tensor<int32, [4]> var_2908 = const()[name = string("op_2908"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_2909 = reshape(shape = var_2908, x = var_2907)[name = string("op_2909")];
            tensor<int32, [4]> var_2910 = const()[name = string("op_2910"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2917_pad_type_0 = const()[name = string("op_2917_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2917_strides_0 = const()[name = string("op_2917_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2917_pad_0 = const()[name = string("op_2917_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2917_dilations_0 = const()[name = string("op_2917_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2917_groups_0 = const()[name = string("op_2917_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2917 = conv(dilations = var_2917_dilations_0, groups = var_2917_groups_0, pad = var_2917_pad_0, pad_type = var_2917_pad_type_0, strides = var_2917_strides_0, weight = encoder_layers_14_self_attn_k_proj_weight, x = input_141)[name = string("op_2917")];
            tensor<int32, [4]> var_2918 = const()[name = string("op_2918"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_2919 = reshape(shape = var_2918, x = var_2917)[name = string("op_2919")];
            tensor<int32, [4]> var_2920 = const()[name = string("op_2920"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2927_pad_type_0 = const()[name = string("op_2927_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2927_strides_0 = const()[name = string("op_2927_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2927_pad_0 = const()[name = string("op_2927_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2927_dilations_0 = const()[name = string("op_2927_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2927_groups_0 = const()[name = string("op_2927_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_2927 = conv(dilations = var_2927_dilations_0, groups = var_2927_groups_0, pad = var_2927_pad_0, pad_type = var_2927_pad_type_0, strides = var_2927_strides_0, weight = encoder_layers_14_self_attn_v_proj_weight, x = input_141)[name = string("op_2927")];
            tensor<int32, [4]> var_2928 = const()[name = string("op_2928"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_2929 = reshape(shape = var_2928, x = var_2927)[name = string("op_2929")];
            tensor<int32, [4]> var_2930 = const()[name = string("op_2930"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_57_to_fp16 = const()[name = string("op_5_promoted_57_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_85 = transpose(perm = var_2910, x = var_2909)[name = string("transpose_124")];
            tensor<fp16, [1, 16, 1024, 128]> var_2936_cast_fp16 = pow(x = q_85, y = var_5_promoted_57_to_fp16)[name = string("op_2936_cast_fp16")];
            tensor<int32, [1]> var_115_axes_0 = const()[name = string("var_115_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_115_keep_dims_0 = const()[name = string("var_115_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_115_cast_fp16 = reduce_mean(axes = var_115_axes_0, keep_dims = var_115_keep_dims_0, x = var_2936_cast_fp16)[name = string("var_115_cast_fp16")];
            fp16 var_2939_to_fp16 = const()[name = string("op_2939_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_2940_cast_fp16 = add(x = var_115_cast_fp16, y = var_2939_to_fp16)[name = string("op_2940_cast_fp16")];
            fp32 var_2941_epsilon_0 = const()[name = string("op_2941_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_2941_cast_fp16 = rsqrt(epsilon = var_2941_epsilon_0, x = var_2940_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_487_cast_fp16 = mul(x = q_85, y = var_2941_cast_fp16)[name = string("x_487_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_87 = mul(x = x_487_cast_fp16, y = encoder_layers_14_self_attn_q_norm_weight)[name = string("q_87")];
            fp16 var_5_promoted_58_to_fp16 = const()[name = string("op_5_promoted_58_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_85 = transpose(perm = var_2920, x = var_2919)[name = string("transpose_123")];
            tensor<fp16, [1, 8, 1024, 128]> var_2949_cast_fp16 = pow(x = k_85, y = var_5_promoted_58_to_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [1]> var_117_axes_0 = const()[name = string("var_117_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_117_keep_dims_0 = const()[name = string("var_117_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_117_cast_fp16 = reduce_mean(axes = var_117_axes_0, keep_dims = var_117_keep_dims_0, x = var_2949_cast_fp16)[name = string("var_117_cast_fp16")];
            fp16 var_2952_to_fp16 = const()[name = string("op_2952_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_2953_cast_fp16 = add(x = var_117_cast_fp16, y = var_2952_to_fp16)[name = string("op_2953_cast_fp16")];
            fp32 var_2954_epsilon_0 = const()[name = string("op_2954_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_2954_cast_fp16 = rsqrt(epsilon = var_2954_epsilon_0, x = var_2953_cast_fp16)[name = string("op_2954_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_493_cast_fp16 = mul(x = k_85, y = var_2954_cast_fp16)[name = string("x_493_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_87 = mul(x = x_493_cast_fp16, y = encoder_layers_14_self_attn_k_norm_weight)[name = string("k_87")];
            tensor<fp16, [1, 16, 1024, 128]> var_2958 = mul(x = q_87, y = cos)[name = string("op_2958")];
            tensor<int32, [2]> var_2959_split_sizes_0 = const()[name = string("op_2959_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2959_axis_0 = const()[name = string("op_2959_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_2959_0, tensor<fp16, [1, 16, 1024, 64]> var_2959_1 = split(axis = var_2959_axis_0, split_sizes = var_2959_split_sizes_0, x = q_87)[name = string("op_2959")];
            fp16 const_45_promoted = const()[name = string("const_45_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_2961 = mul(x = var_2959_1, y = const_45_promoted)[name = string("op_2961")];
            bool var_2963_interleave_0 = const()[name = string("op_2963_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_2963 = concat(axis = var_17, interleave = var_2963_interleave_0, values = (var_2961, var_2959_0))[name = string("op_2963")];
            tensor<fp16, [1, 16, 1024, 128]> var_2964 = mul(x = var_2963, y = sin)[name = string("op_2964")];
            tensor<fp16, [1, 16, 1024, 128]> query_29 = add(x = var_2958, y = var_2964)[name = string("query_29")];
            tensor<fp16, [1, 8, 1024, 128]> var_2966 = mul(x = k_87, y = cos)[name = string("op_2966")];
            tensor<int32, [2]> var_2967_split_sizes_0 = const()[name = string("op_2967_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2967_axis_0 = const()[name = string("op_2967_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_2967_0, tensor<fp16, [1, 8, 1024, 64]> var_2967_1 = split(axis = var_2967_axis_0, split_sizes = var_2967_split_sizes_0, x = k_87)[name = string("op_2967")];
            fp16 const_46_promoted = const()[name = string("const_46_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_2969 = mul(x = var_2967_1, y = const_46_promoted)[name = string("op_2969")];
            bool var_2971_interleave_0 = const()[name = string("op_2971_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_2971 = concat(axis = var_17, interleave = var_2971_interleave_0, values = (var_2969, var_2967_0))[name = string("op_2971")];
            tensor<fp16, [1, 8, 1024, 128]> var_2972 = mul(x = var_2971, y = sin)[name = string("op_2972")];
            tensor<fp16, [1, 8, 1024, 128]> x_495 = add(x = var_2966, y = var_2972)[name = string("x_495")];
            tensor<int32, [1]> var_2974_axes_0 = const()[name = string("op_2974_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2974 = expand_dims(axes = var_2974_axes_0, x = x_495)[name = string("op_2974")];
            tensor<int32, [5]> x_497_reps_0 = const()[name = string("x_497_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_497 = tile(reps = x_497_reps_0, x = var_2974)[name = string("x_497")];
            tensor<int32, [4]> var_2977 = const()[name = string("op_2977"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_29 = reshape(shape = var_2977, x = x_497)[name = string("key_29")];
            tensor<int32, [1]> var_2979_axes_0 = const()[name = string("op_2979_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_499 = transpose(perm = var_2930, x = var_2929)[name = string("transpose_122")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2979 = expand_dims(axes = var_2979_axes_0, x = x_499)[name = string("op_2979")];
            tensor<int32, [5]> x_501_reps_0 = const()[name = string("x_501_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_501 = tile(reps = x_501_reps_0, x = var_2979)[name = string("x_501")];
            tensor<int32, [4]> var_2982 = const()[name = string("op_2982"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_29 = reshape(shape = var_2982, x = x_501)[name = string("value_29")];
            bool var_2987_transpose_x_1 = const()[name = string("op_2987_transpose_x_1"), val = bool(false)];
            bool var_2987_transpose_y_1 = const()[name = string("op_2987_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_2987_cast_fp16 = matmul(transpose_x = var_2987_transpose_x_1, transpose_y = var_2987_transpose_y_1, x = query_29, y = key_29)[name = string("op_2987_cast_fp16")];
            fp16 var_2988_to_fp16 = const()[name = string("op_2988_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_85_cast_fp16 = mul(x = var_2987_cast_fp16, y = var_2988_to_fp16)[name = string("attn_weights_85_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_87_cast_fp16 = add(x = attn_weights_85_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_2992_cast_fp16 = softmax(axis = var_17, x = attn_weights_87_cast_fp16)[name = string("op_2992_cast_fp16")];
            bool var_2996_transpose_x_0 = const()[name = string("op_2996_transpose_x_0"), val = bool(false)];
            bool var_2996_transpose_y_0 = const()[name = string("op_2996_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_2996_cast_fp16 = matmul(transpose_x = var_2996_transpose_x_0, transpose_y = var_2996_transpose_y_0, x = var_2992_cast_fp16, y = value_29)[name = string("op_2996_cast_fp16")];
            tensor<int32, [4]> var_2998 = const()[name = string("op_2998"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3001 = const()[name = string("op_3001"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_2999 = transpose(perm = var_2998, x = var_2996_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 1024, 2048]> attn_out_87 = reshape(shape = var_3001, x = var_2999)[name = string("attn_out_87")];
            tensor<int32, [3]> var_3003 = const()[name = string("op_3003"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_14 = const()[name = string("squeeze_14"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1150116800)))];
            string var_3012_pad_type_0 = const()[name = string("op_3012_pad_type_0"), val = string("valid")];
            int32 var_3012_groups_0 = const()[name = string("op_3012_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3012_strides_0 = const()[name = string("op_3012_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3012_pad_0 = const()[name = string("op_3012_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3012_dilations_0 = const()[name = string("op_3012_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_3004 = transpose(perm = var_3003, x = attn_out_87)[name = string("transpose_120")];
            tensor<fp16, [1, 1024, 1024]> var_3012 = conv(dilations = var_3012_dilations_0, groups = var_3012_groups_0, pad = var_3012_pad_0, pad_type = var_3012_pad_type_0, strides = var_3012_strides_0, weight = squeeze_14, x = var_3004)[name = string("op_3012")];
            tensor<int32, [3]> var_3013 = const()[name = string("op_3013"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_89 = transpose(perm = var_3013, x = var_3012)[name = string("transpose_119")];
            tensor<fp16, [1, 1024, 1024]> x_503_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = attn_out_89)[name = string("x_503_cast_fp16")];
            fp16 var_5_promoted_59_to_fp16 = const()[name = string("op_5_promoted_59_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_3019_cast_fp16 = pow(x = x_503_cast_fp16, y = var_5_promoted_59_to_fp16)[name = string("op_3019_cast_fp16")];
            tensor<int32, [1]> var_119_axes_0 = const()[name = string("var_119_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_119_keep_dims_0 = const()[name = string("var_119_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_119_cast_fp16 = reduce_mean(axes = var_119_axes_0, keep_dims = var_119_keep_dims_0, x = var_3019_cast_fp16)[name = string("var_119_cast_fp16")];
            fp16 var_3022_to_fp16 = const()[name = string("op_3022_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3023_cast_fp16 = add(x = var_119_cast_fp16, y = var_3022_to_fp16)[name = string("op_3023_cast_fp16")];
            fp32 var_3024_epsilon_0 = const()[name = string("op_3024_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3024_cast_fp16 = rsqrt(epsilon = var_3024_epsilon_0, x = var_3023_cast_fp16)[name = string("op_3024_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_507_cast_fp16 = mul(x = x_503_cast_fp16, y = var_3024_cast_fp16)[name = string("x_507_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_14_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_14_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1154311168)))];
            tensor<fp16, [1, 1024, 1024]> var_3027_cast_fp16 = mul(x = x_507_cast_fp16, y = encoder_layers_14_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_3027_cast_fp16")];
            tensor<int32, [3]> var_3032 = const()[name = string("op_3032"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_145_axes_0 = const()[name = string("input_145_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3033 = transpose(perm = var_3032, x = var_3027_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 1024, 1, 1024]> input_145 = expand_dims(axes = input_145_axes_0, x = var_3033)[name = string("input_145")];
            string input_147_pad_type_0 = const()[name = string("input_147_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_147_strides_0 = const()[name = string("input_147_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_147_pad_0 = const()[name = string("input_147_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_147_dilations_0 = const()[name = string("input_147_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_147_groups_0 = const()[name = string("input_147_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_147 = conv(dilations = input_147_dilations_0, groups = input_147_groups_0, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = input_147_strides_0, weight = encoder_layers_14_mlp_gate_proj_weight, x = input_145)[name = string("input_147")];
            string up_29_pad_type_0 = const()[name = string("up_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_29_strides_0 = const()[name = string("up_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_29_pad_0 = const()[name = string("up_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_29_dilations_0 = const()[name = string("up_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_29_groups_0 = const()[name = string("up_29_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_29 = conv(dilations = up_29_dilations_0, groups = up_29_groups_0, pad = up_29_pad_0, pad_type = up_29_pad_type_0, strides = up_29_strides_0, weight = encoder_layers_14_mlp_up_proj_weight, x = input_145)[name = string("up_29")];
            tensor<fp16, [1, 3072, 1, 1024]> var_3047 = silu(x = input_147)[name = string("op_3047")];
            tensor<fp16, [1, 3072, 1, 1024]> input_149 = mul(x = var_3047, y = up_29)[name = string("input_149")];
            string var_3054_pad_type_0 = const()[name = string("op_3054_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3054_strides_0 = const()[name = string("op_3054_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3054_pad_0 = const()[name = string("op_3054_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3054_dilations_0 = const()[name = string("op_3054_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3054_groups_0 = const()[name = string("op_3054_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3054 = conv(dilations = var_3054_dilations_0, groups = var_3054_groups_0, pad = var_3054_pad_0, pad_type = var_3054_pad_type_0, strides = var_3054_strides_0, weight = encoder_layers_14_mlp_down_proj_weight, x = input_149)[name = string("op_3054")];
            tensor<int32, [1]> var_3055_axes_0 = const()[name = string("op_3055_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3055 = squeeze(axes = var_3055_axes_0, x = var_3054)[name = string("op_3055")];
            tensor<int32, [3]> var_3056 = const()[name = string("op_3056"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_29 = transpose(perm = var_3056, x = var_3055)[name = string("transpose_117")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_31_cast_fp16 = add(x = x_503_cast_fp16, y = mlp_out_29)[name = string("hidden_states_31_cast_fp16")];
            fp16 var_5_promoted_60_to_fp16 = const()[name = string("op_5_promoted_60_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_3083_cast_fp16 = pow(x = hidden_states_31_cast_fp16, y = var_5_promoted_60_to_fp16)[name = string("op_3083_cast_fp16")];
            tensor<int32, [1]> var_121_axes_0 = const()[name = string("var_121_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_121_keep_dims_0 = const()[name = string("var_121_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_121_cast_fp16 = reduce_mean(axes = var_121_axes_0, keep_dims = var_121_keep_dims_0, x = var_3083_cast_fp16)[name = string("var_121_cast_fp16")];
            fp16 var_3086_to_fp16 = const()[name = string("op_3086_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3087_cast_fp16 = add(x = var_121_cast_fp16, y = var_3086_to_fp16)[name = string("op_3087_cast_fp16")];
            fp32 var_3088_epsilon_0 = const()[name = string("op_3088_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3088_cast_fp16 = rsqrt(epsilon = var_3088_epsilon_0, x = var_3087_cast_fp16)[name = string("op_3088_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_513_cast_fp16 = mul(x = hidden_states_31_cast_fp16, y = var_3088_cast_fp16)[name = string("x_513_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_15_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_15_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1154313280)))];
            tensor<fp16, [1, 1024, 1024]> var_3091_cast_fp16 = mul(x = x_513_cast_fp16, y = encoder_layers_15_input_layernorm_weight_promoted_to_fp16)[name = string("op_3091_cast_fp16")];
            tensor<int32, [3]> var_3096 = const()[name = string("op_3096"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_151_axes_0 = const()[name = string("input_151_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3097 = transpose(perm = var_3096, x = var_3091_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 1024, 1, 1024]> input_151 = expand_dims(axes = input_151_axes_0, x = var_3097)[name = string("input_151")];
            string var_3104_pad_type_0 = const()[name = string("op_3104_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3104_strides_0 = const()[name = string("op_3104_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3104_pad_0 = const()[name = string("op_3104_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3104_dilations_0 = const()[name = string("op_3104_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3104_groups_0 = const()[name = string("op_3104_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_3104 = conv(dilations = var_3104_dilations_0, groups = var_3104_groups_0, pad = var_3104_pad_0, pad_type = var_3104_pad_type_0, strides = var_3104_strides_0, weight = encoder_layers_15_self_attn_q_proj_weight, x = input_151)[name = string("op_3104")];
            tensor<int32, [4]> var_3105 = const()[name = string("op_3105"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_3106 = reshape(shape = var_3105, x = var_3104)[name = string("op_3106")];
            tensor<int32, [4]> var_3107 = const()[name = string("op_3107"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3114_pad_type_0 = const()[name = string("op_3114_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3114_strides_0 = const()[name = string("op_3114_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3114_pad_0 = const()[name = string("op_3114_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3114_dilations_0 = const()[name = string("op_3114_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3114_groups_0 = const()[name = string("op_3114_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3114 = conv(dilations = var_3114_dilations_0, groups = var_3114_groups_0, pad = var_3114_pad_0, pad_type = var_3114_pad_type_0, strides = var_3114_strides_0, weight = encoder_layers_15_self_attn_k_proj_weight, x = input_151)[name = string("op_3114")];
            tensor<int32, [4]> var_3115 = const()[name = string("op_3115"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_3116 = reshape(shape = var_3115, x = var_3114)[name = string("op_3116")];
            tensor<int32, [4]> var_3117 = const()[name = string("op_3117"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3124_pad_type_0 = const()[name = string("op_3124_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3124_strides_0 = const()[name = string("op_3124_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3124_pad_0 = const()[name = string("op_3124_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3124_dilations_0 = const()[name = string("op_3124_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3124_groups_0 = const()[name = string("op_3124_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3124 = conv(dilations = var_3124_dilations_0, groups = var_3124_groups_0, pad = var_3124_pad_0, pad_type = var_3124_pad_type_0, strides = var_3124_strides_0, weight = encoder_layers_15_self_attn_v_proj_weight, x = input_151)[name = string("op_3124")];
            tensor<int32, [4]> var_3125 = const()[name = string("op_3125"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_3126 = reshape(shape = var_3125, x = var_3124)[name = string("op_3126")];
            tensor<int32, [4]> var_3127 = const()[name = string("op_3127"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_61_to_fp16 = const()[name = string("op_5_promoted_61_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_91 = transpose(perm = var_3107, x = var_3106)[name = string("transpose_115")];
            tensor<fp16, [1, 16, 1024, 128]> var_3133_cast_fp16 = pow(x = q_91, y = var_5_promoted_61_to_fp16)[name = string("op_3133_cast_fp16")];
            tensor<int32, [1]> var_123_axes_0 = const()[name = string("var_123_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_123_keep_dims_0 = const()[name = string("var_123_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_123_cast_fp16 = reduce_mean(axes = var_123_axes_0, keep_dims = var_123_keep_dims_0, x = var_3133_cast_fp16)[name = string("var_123_cast_fp16")];
            fp16 var_3136_to_fp16 = const()[name = string("op_3136_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_3137_cast_fp16 = add(x = var_123_cast_fp16, y = var_3136_to_fp16)[name = string("op_3137_cast_fp16")];
            fp32 var_3138_epsilon_0 = const()[name = string("op_3138_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_3138_cast_fp16 = rsqrt(epsilon = var_3138_epsilon_0, x = var_3137_cast_fp16)[name = string("op_3138_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_521_cast_fp16 = mul(x = q_91, y = var_3138_cast_fp16)[name = string("x_521_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_93 = mul(x = x_521_cast_fp16, y = encoder_layers_15_self_attn_q_norm_weight)[name = string("q_93")];
            fp16 var_5_promoted_62_to_fp16 = const()[name = string("op_5_promoted_62_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_91 = transpose(perm = var_3117, x = var_3116)[name = string("transpose_114")];
            tensor<fp16, [1, 8, 1024, 128]> var_3146_cast_fp16 = pow(x = k_91, y = var_5_promoted_62_to_fp16)[name = string("op_3146_cast_fp16")];
            tensor<int32, [1]> var_125_axes_0 = const()[name = string("var_125_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_125_keep_dims_0 = const()[name = string("var_125_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_125_cast_fp16 = reduce_mean(axes = var_125_axes_0, keep_dims = var_125_keep_dims_0, x = var_3146_cast_fp16)[name = string("var_125_cast_fp16")];
            fp16 var_3149_to_fp16 = const()[name = string("op_3149_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_3150_cast_fp16 = add(x = var_125_cast_fp16, y = var_3149_to_fp16)[name = string("op_3150_cast_fp16")];
            fp32 var_3151_epsilon_0 = const()[name = string("op_3151_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_3151_cast_fp16 = rsqrt(epsilon = var_3151_epsilon_0, x = var_3150_cast_fp16)[name = string("op_3151_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_527_cast_fp16 = mul(x = k_91, y = var_3151_cast_fp16)[name = string("x_527_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_93 = mul(x = x_527_cast_fp16, y = encoder_layers_15_self_attn_k_norm_weight)[name = string("k_93")];
            tensor<fp16, [1, 16, 1024, 128]> var_3155 = mul(x = q_93, y = cos)[name = string("op_3155")];
            tensor<int32, [2]> var_3156_split_sizes_0 = const()[name = string("op_3156_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3156_axis_0 = const()[name = string("op_3156_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_3156_0, tensor<fp16, [1, 16, 1024, 64]> var_3156_1 = split(axis = var_3156_axis_0, split_sizes = var_3156_split_sizes_0, x = q_93)[name = string("op_3156")];
            fp16 const_48_promoted = const()[name = string("const_48_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_3158 = mul(x = var_3156_1, y = const_48_promoted)[name = string("op_3158")];
            bool var_3160_interleave_0 = const()[name = string("op_3160_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_3160 = concat(axis = var_17, interleave = var_3160_interleave_0, values = (var_3158, var_3156_0))[name = string("op_3160")];
            tensor<fp16, [1, 16, 1024, 128]> var_3161 = mul(x = var_3160, y = sin)[name = string("op_3161")];
            tensor<fp16, [1, 16, 1024, 128]> query_31 = add(x = var_3155, y = var_3161)[name = string("query_31")];
            tensor<fp16, [1, 8, 1024, 128]> var_3163 = mul(x = k_93, y = cos)[name = string("op_3163")];
            tensor<int32, [2]> var_3164_split_sizes_0 = const()[name = string("op_3164_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3164_axis_0 = const()[name = string("op_3164_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_3164_0, tensor<fp16, [1, 8, 1024, 64]> var_3164_1 = split(axis = var_3164_axis_0, split_sizes = var_3164_split_sizes_0, x = k_93)[name = string("op_3164")];
            fp16 const_49_promoted = const()[name = string("const_49_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_3166 = mul(x = var_3164_1, y = const_49_promoted)[name = string("op_3166")];
            bool var_3168_interleave_0 = const()[name = string("op_3168_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_3168 = concat(axis = var_17, interleave = var_3168_interleave_0, values = (var_3166, var_3164_0))[name = string("op_3168")];
            tensor<fp16, [1, 8, 1024, 128]> var_3169 = mul(x = var_3168, y = sin)[name = string("op_3169")];
            tensor<fp16, [1, 8, 1024, 128]> x_529 = add(x = var_3163, y = var_3169)[name = string("x_529")];
            tensor<int32, [1]> var_3171_axes_0 = const()[name = string("op_3171_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3171 = expand_dims(axes = var_3171_axes_0, x = x_529)[name = string("op_3171")];
            tensor<int32, [5]> x_531_reps_0 = const()[name = string("x_531_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_531 = tile(reps = x_531_reps_0, x = var_3171)[name = string("x_531")];
            tensor<int32, [4]> var_3174 = const()[name = string("op_3174"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_31 = reshape(shape = var_3174, x = x_531)[name = string("key_31")];
            tensor<int32, [1]> var_3176_axes_0 = const()[name = string("op_3176_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_533 = transpose(perm = var_3127, x = var_3126)[name = string("transpose_113")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3176 = expand_dims(axes = var_3176_axes_0, x = x_533)[name = string("op_3176")];
            tensor<int32, [5]> x_535_reps_0 = const()[name = string("x_535_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_535 = tile(reps = x_535_reps_0, x = var_3176)[name = string("x_535")];
            tensor<int32, [4]> var_3179 = const()[name = string("op_3179"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_31 = reshape(shape = var_3179, x = x_535)[name = string("value_31")];
            bool var_3184_transpose_x_1 = const()[name = string("op_3184_transpose_x_1"), val = bool(false)];
            bool var_3184_transpose_y_1 = const()[name = string("op_3184_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_3184_cast_fp16 = matmul(transpose_x = var_3184_transpose_x_1, transpose_y = var_3184_transpose_y_1, x = query_31, y = key_31)[name = string("op_3184_cast_fp16")];
            fp16 var_3185_to_fp16 = const()[name = string("op_3185_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_91_cast_fp16 = mul(x = var_3184_cast_fp16, y = var_3185_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_3189_cast_fp16 = softmax(axis = var_17, x = attn_weights_93_cast_fp16)[name = string("op_3189_cast_fp16")];
            bool var_3193_transpose_x_0 = const()[name = string("op_3193_transpose_x_0"), val = bool(false)];
            bool var_3193_transpose_y_0 = const()[name = string("op_3193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_3193_cast_fp16 = matmul(transpose_x = var_3193_transpose_x_0, transpose_y = var_3193_transpose_y_0, x = var_3189_cast_fp16, y = value_31)[name = string("op_3193_cast_fp16")];
            tensor<int32, [4]> var_3195 = const()[name = string("op_3195"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3198 = const()[name = string("op_3198"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_3196 = transpose(perm = var_3195, x = var_3193_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 1024, 2048]> attn_out_93 = reshape(shape = var_3198, x = var_3196)[name = string("attn_out_93")];
            tensor<int32, [3]> var_3200 = const()[name = string("op_3200"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_15 = const()[name = string("squeeze_15"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1154315392)))];
            string var_3209_pad_type_0 = const()[name = string("op_3209_pad_type_0"), val = string("valid")];
            int32 var_3209_groups_0 = const()[name = string("op_3209_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3209_strides_0 = const()[name = string("op_3209_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3209_pad_0 = const()[name = string("op_3209_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3209_dilations_0 = const()[name = string("op_3209_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_3201 = transpose(perm = var_3200, x = attn_out_93)[name = string("transpose_111")];
            tensor<fp16, [1, 1024, 1024]> var_3209 = conv(dilations = var_3209_dilations_0, groups = var_3209_groups_0, pad = var_3209_pad_0, pad_type = var_3209_pad_type_0, strides = var_3209_strides_0, weight = squeeze_15, x = var_3201)[name = string("op_3209")];
            tensor<int32, [3]> var_3210 = const()[name = string("op_3210"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_95 = transpose(perm = var_3210, x = var_3209)[name = string("transpose_110")];
            tensor<fp16, [1, 1024, 1024]> x_537_cast_fp16 = add(x = hidden_states_31_cast_fp16, y = attn_out_95)[name = string("x_537_cast_fp16")];
            fp16 var_5_promoted_63_to_fp16 = const()[name = string("op_5_promoted_63_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_3216_cast_fp16 = pow(x = x_537_cast_fp16, y = var_5_promoted_63_to_fp16)[name = string("op_3216_cast_fp16")];
            tensor<int32, [1]> var_127_axes_0 = const()[name = string("var_127_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_127_keep_dims_0 = const()[name = string("var_127_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_127_cast_fp16 = reduce_mean(axes = var_127_axes_0, keep_dims = var_127_keep_dims_0, x = var_3216_cast_fp16)[name = string("var_127_cast_fp16")];
            fp16 var_3219_to_fp16 = const()[name = string("op_3219_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3220_cast_fp16 = add(x = var_127_cast_fp16, y = var_3219_to_fp16)[name = string("op_3220_cast_fp16")];
            fp32 var_3221_epsilon_0 = const()[name = string("op_3221_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3221_cast_fp16 = rsqrt(epsilon = var_3221_epsilon_0, x = var_3220_cast_fp16)[name = string("op_3221_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_541_cast_fp16 = mul(x = x_537_cast_fp16, y = var_3221_cast_fp16)[name = string("x_541_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_15_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_15_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1158509760)))];
            tensor<fp16, [1, 1024, 1024]> var_3224_cast_fp16 = mul(x = x_541_cast_fp16, y = encoder_layers_15_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_3224_cast_fp16")];
            tensor<int32, [3]> var_3229 = const()[name = string("op_3229"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_155_axes_0 = const()[name = string("input_155_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3230 = transpose(perm = var_3229, x = var_3224_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 1024, 1, 1024]> input_155 = expand_dims(axes = input_155_axes_0, x = var_3230)[name = string("input_155")];
            string input_157_pad_type_0 = const()[name = string("input_157_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_157_strides_0 = const()[name = string("input_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_157_pad_0 = const()[name = string("input_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_157_dilations_0 = const()[name = string("input_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_157_groups_0 = const()[name = string("input_157_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_157 = conv(dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = encoder_layers_15_mlp_gate_proj_weight, x = input_155)[name = string("input_157")];
            string up_31_pad_type_0 = const()[name = string("up_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_31_strides_0 = const()[name = string("up_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_31_pad_0 = const()[name = string("up_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_31_dilations_0 = const()[name = string("up_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_31_groups_0 = const()[name = string("up_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_31 = conv(dilations = up_31_dilations_0, groups = up_31_groups_0, pad = up_31_pad_0, pad_type = up_31_pad_type_0, strides = up_31_strides_0, weight = encoder_layers_15_mlp_up_proj_weight, x = input_155)[name = string("up_31")];
            tensor<fp16, [1, 3072, 1, 1024]> var_3244 = silu(x = input_157)[name = string("op_3244")];
            tensor<fp16, [1, 3072, 1, 1024]> input_159 = mul(x = var_3244, y = up_31)[name = string("input_159")];
            string var_3251_pad_type_0 = const()[name = string("op_3251_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3251_strides_0 = const()[name = string("op_3251_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3251_pad_0 = const()[name = string("op_3251_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3251_dilations_0 = const()[name = string("op_3251_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3251_groups_0 = const()[name = string("op_3251_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3251 = conv(dilations = var_3251_dilations_0, groups = var_3251_groups_0, pad = var_3251_pad_0, pad_type = var_3251_pad_type_0, strides = var_3251_strides_0, weight = encoder_layers_15_mlp_down_proj_weight, x = input_159)[name = string("op_3251")];
            tensor<int32, [1]> var_3252_axes_0 = const()[name = string("op_3252_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3252 = squeeze(axes = var_3252_axes_0, x = var_3251)[name = string("op_3252")];
            tensor<int32, [3]> var_3253 = const()[name = string("op_3253"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_31 = transpose(perm = var_3253, x = var_3252)[name = string("transpose_108")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_33_cast_fp16 = add(x = x_537_cast_fp16, y = mlp_out_31)[name = string("hidden_states_33_cast_fp16")];
            fp16 var_5_promoted_64_to_fp16 = const()[name = string("op_5_promoted_64_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_3280_cast_fp16 = pow(x = hidden_states_33_cast_fp16, y = var_5_promoted_64_to_fp16)[name = string("op_3280_cast_fp16")];
            tensor<int32, [1]> var_129_axes_0 = const()[name = string("var_129_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_129_keep_dims_0 = const()[name = string("var_129_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_129_cast_fp16 = reduce_mean(axes = var_129_axes_0, keep_dims = var_129_keep_dims_0, x = var_3280_cast_fp16)[name = string("var_129_cast_fp16")];
            fp16 var_3283_to_fp16 = const()[name = string("op_3283_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3284_cast_fp16 = add(x = var_129_cast_fp16, y = var_3283_to_fp16)[name = string("op_3284_cast_fp16")];
            fp32 var_3285_epsilon_0 = const()[name = string("op_3285_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3285_cast_fp16 = rsqrt(epsilon = var_3285_epsilon_0, x = var_3284_cast_fp16)[name = string("op_3285_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_547_cast_fp16 = mul(x = hidden_states_33_cast_fp16, y = var_3285_cast_fp16)[name = string("x_547_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_16_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_16_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1158511872)))];
            tensor<fp16, [1, 1024, 1024]> var_3288_cast_fp16 = mul(x = x_547_cast_fp16, y = encoder_layers_16_input_layernorm_weight_promoted_to_fp16)[name = string("op_3288_cast_fp16")];
            tensor<int32, [3]> var_3293 = const()[name = string("op_3293"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_161_axes_0 = const()[name = string("input_161_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3294 = transpose(perm = var_3293, x = var_3288_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 1024, 1, 1024]> input_161 = expand_dims(axes = input_161_axes_0, x = var_3294)[name = string("input_161")];
            string var_3301_pad_type_0 = const()[name = string("op_3301_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3301_strides_0 = const()[name = string("op_3301_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3301_pad_0 = const()[name = string("op_3301_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3301_dilations_0 = const()[name = string("op_3301_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3301_groups_0 = const()[name = string("op_3301_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_3301 = conv(dilations = var_3301_dilations_0, groups = var_3301_groups_0, pad = var_3301_pad_0, pad_type = var_3301_pad_type_0, strides = var_3301_strides_0, weight = encoder_layers_16_self_attn_q_proj_weight, x = input_161)[name = string("op_3301")];
            tensor<int32, [4]> var_3302 = const()[name = string("op_3302"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_3303 = reshape(shape = var_3302, x = var_3301)[name = string("op_3303")];
            tensor<int32, [4]> var_3304 = const()[name = string("op_3304"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3311_pad_type_0 = const()[name = string("op_3311_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3311_strides_0 = const()[name = string("op_3311_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3311_pad_0 = const()[name = string("op_3311_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3311_dilations_0 = const()[name = string("op_3311_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3311_groups_0 = const()[name = string("op_3311_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3311 = conv(dilations = var_3311_dilations_0, groups = var_3311_groups_0, pad = var_3311_pad_0, pad_type = var_3311_pad_type_0, strides = var_3311_strides_0, weight = encoder_layers_16_self_attn_k_proj_weight, x = input_161)[name = string("op_3311")];
            tensor<int32, [4]> var_3312 = const()[name = string("op_3312"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_3313 = reshape(shape = var_3312, x = var_3311)[name = string("op_3313")];
            tensor<int32, [4]> var_3314 = const()[name = string("op_3314"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3321_pad_type_0 = const()[name = string("op_3321_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3321_strides_0 = const()[name = string("op_3321_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3321_pad_0 = const()[name = string("op_3321_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3321_dilations_0 = const()[name = string("op_3321_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3321_groups_0 = const()[name = string("op_3321_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3321 = conv(dilations = var_3321_dilations_0, groups = var_3321_groups_0, pad = var_3321_pad_0, pad_type = var_3321_pad_type_0, strides = var_3321_strides_0, weight = encoder_layers_16_self_attn_v_proj_weight, x = input_161)[name = string("op_3321")];
            tensor<int32, [4]> var_3322 = const()[name = string("op_3322"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_3323 = reshape(shape = var_3322, x = var_3321)[name = string("op_3323")];
            tensor<int32, [4]> var_3324 = const()[name = string("op_3324"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_65_to_fp16 = const()[name = string("op_5_promoted_65_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_97 = transpose(perm = var_3304, x = var_3303)[name = string("transpose_106")];
            tensor<fp16, [1, 16, 1024, 128]> var_3330_cast_fp16 = pow(x = q_97, y = var_5_promoted_65_to_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [1]> var_131_axes_0 = const()[name = string("var_131_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_131_keep_dims_0 = const()[name = string("var_131_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_131_cast_fp16 = reduce_mean(axes = var_131_axes_0, keep_dims = var_131_keep_dims_0, x = var_3330_cast_fp16)[name = string("var_131_cast_fp16")];
            fp16 var_3333_to_fp16 = const()[name = string("op_3333_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_3334_cast_fp16 = add(x = var_131_cast_fp16, y = var_3333_to_fp16)[name = string("op_3334_cast_fp16")];
            fp32 var_3335_epsilon_0 = const()[name = string("op_3335_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_3335_cast_fp16 = rsqrt(epsilon = var_3335_epsilon_0, x = var_3334_cast_fp16)[name = string("op_3335_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_555_cast_fp16 = mul(x = q_97, y = var_3335_cast_fp16)[name = string("x_555_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_99 = mul(x = x_555_cast_fp16, y = encoder_layers_16_self_attn_q_norm_weight)[name = string("q_99")];
            fp16 var_5_promoted_66_to_fp16 = const()[name = string("op_5_promoted_66_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_97 = transpose(perm = var_3314, x = var_3313)[name = string("transpose_105")];
            tensor<fp16, [1, 8, 1024, 128]> var_3343_cast_fp16 = pow(x = k_97, y = var_5_promoted_66_to_fp16)[name = string("op_3343_cast_fp16")];
            tensor<int32, [1]> var_133_axes_0 = const()[name = string("var_133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_133_keep_dims_0 = const()[name = string("var_133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_133_cast_fp16_0 = reduce_mean(axes = var_133_axes_0, keep_dims = var_133_keep_dims_0, x = var_3343_cast_fp16)[name = string("var_133_cast_fp16")];
            fp16 var_3346_to_fp16 = const()[name = string("op_3346_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_3347_cast_fp16 = add(x = var_133_cast_fp16_0, y = var_3346_to_fp16)[name = string("op_3347_cast_fp16")];
            fp32 var_3348_epsilon_0 = const()[name = string("op_3348_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_3348_cast_fp16 = rsqrt(epsilon = var_3348_epsilon_0, x = var_3347_cast_fp16)[name = string("op_3348_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_561_cast_fp16 = mul(x = k_97, y = var_3348_cast_fp16)[name = string("x_561_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_99 = mul(x = x_561_cast_fp16, y = encoder_layers_16_self_attn_k_norm_weight)[name = string("k_99")];
            tensor<fp16, [1, 16, 1024, 128]> var_3352 = mul(x = q_99, y = cos)[name = string("op_3352")];
            tensor<int32, [2]> var_3353_split_sizes_0 = const()[name = string("op_3353_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3353_axis_0 = const()[name = string("op_3353_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_3353_0, tensor<fp16, [1, 16, 1024, 64]> var_3353_1 = split(axis = var_3353_axis_0, split_sizes = var_3353_split_sizes_0, x = q_99)[name = string("op_3353")];
            fp16 const_51_promoted = const()[name = string("const_51_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_3355 = mul(x = var_3353_1, y = const_51_promoted)[name = string("op_3355")];
            bool var_3357_interleave_0 = const()[name = string("op_3357_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_3357 = concat(axis = var_17, interleave = var_3357_interleave_0, values = (var_3355, var_3353_0))[name = string("op_3357")];
            tensor<fp16, [1, 16, 1024, 128]> var_3358 = mul(x = var_3357, y = sin)[name = string("op_3358")];
            tensor<fp16, [1, 16, 1024, 128]> query_33 = add(x = var_3352, y = var_3358)[name = string("query_33")];
            tensor<fp16, [1, 8, 1024, 128]> var_3360 = mul(x = k_99, y = cos)[name = string("op_3360")];
            tensor<int32, [2]> var_3361_split_sizes_0 = const()[name = string("op_3361_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3361_axis_0 = const()[name = string("op_3361_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_3361_0, tensor<fp16, [1, 8, 1024, 64]> var_3361_1 = split(axis = var_3361_axis_0, split_sizes = var_3361_split_sizes_0, x = k_99)[name = string("op_3361")];
            fp16 const_52_promoted = const()[name = string("const_52_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_3363 = mul(x = var_3361_1, y = const_52_promoted)[name = string("op_3363")];
            bool var_3365_interleave_0 = const()[name = string("op_3365_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_3365 = concat(axis = var_17, interleave = var_3365_interleave_0, values = (var_3363, var_3361_0))[name = string("op_3365")];
            tensor<fp16, [1, 8, 1024, 128]> var_3366 = mul(x = var_3365, y = sin)[name = string("op_3366")];
            tensor<fp16, [1, 8, 1024, 128]> x_563 = add(x = var_3360, y = var_3366)[name = string("x_563")];
            tensor<int32, [1]> var_3368_axes_0 = const()[name = string("op_3368_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3368 = expand_dims(axes = var_3368_axes_0, x = x_563)[name = string("op_3368")];
            tensor<int32, [5]> x_565_reps_0 = const()[name = string("x_565_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_565 = tile(reps = x_565_reps_0, x = var_3368)[name = string("x_565")];
            tensor<int32, [4]> var_3371 = const()[name = string("op_3371"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_33 = reshape(shape = var_3371, x = x_565)[name = string("key_33")];
            tensor<int32, [1]> var_3373_axes_0 = const()[name = string("op_3373_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_567 = transpose(perm = var_3324, x = var_3323)[name = string("transpose_104")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3373 = expand_dims(axes = var_3373_axes_0, x = x_567)[name = string("op_3373")];
            tensor<int32, [5]> x_569_reps_0 = const()[name = string("x_569_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_569 = tile(reps = x_569_reps_0, x = var_3373)[name = string("x_569")];
            tensor<int32, [4]> var_3376 = const()[name = string("op_3376"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_33 = reshape(shape = var_3376, x = x_569)[name = string("value_33")];
            bool var_3381_transpose_x_1 = const()[name = string("op_3381_transpose_x_1"), val = bool(false)];
            bool var_3381_transpose_y_1 = const()[name = string("op_3381_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_3381_cast_fp16 = matmul(transpose_x = var_3381_transpose_x_1, transpose_y = var_3381_transpose_y_1, x = query_33, y = key_33)[name = string("op_3381_cast_fp16")];
            fp16 var_3382_to_fp16 = const()[name = string("op_3382_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_97_cast_fp16 = mul(x = var_3381_cast_fp16, y = var_3382_to_fp16)[name = string("attn_weights_97_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_99_cast_fp16 = add(x = attn_weights_97_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_3386_cast_fp16 = softmax(axis = var_17, x = attn_weights_99_cast_fp16)[name = string("op_3386_cast_fp16")];
            bool var_3390_transpose_x_0 = const()[name = string("op_3390_transpose_x_0"), val = bool(false)];
            bool var_3390_transpose_y_0 = const()[name = string("op_3390_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_3390_cast_fp16 = matmul(transpose_x = var_3390_transpose_x_0, transpose_y = var_3390_transpose_y_0, x = var_3386_cast_fp16, y = value_33)[name = string("op_3390_cast_fp16")];
            tensor<int32, [4]> var_3392 = const()[name = string("op_3392"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3395 = const()[name = string("op_3395"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_3393 = transpose(perm = var_3392, x = var_3390_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 1024, 2048]> attn_out_99 = reshape(shape = var_3395, x = var_3393)[name = string("attn_out_99")];
            tensor<int32, [3]> var_3397 = const()[name = string("op_3397"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_16 = const()[name = string("squeeze_16"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1158513984)))];
            string var_3406_pad_type_0 = const()[name = string("op_3406_pad_type_0"), val = string("valid")];
            int32 var_3406_groups_0 = const()[name = string("op_3406_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3406_strides_0 = const()[name = string("op_3406_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3406_pad_0 = const()[name = string("op_3406_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3406_dilations_0 = const()[name = string("op_3406_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_3398 = transpose(perm = var_3397, x = attn_out_99)[name = string("transpose_102")];
            tensor<fp16, [1, 1024, 1024]> var_3406 = conv(dilations = var_3406_dilations_0, groups = var_3406_groups_0, pad = var_3406_pad_0, pad_type = var_3406_pad_type_0, strides = var_3406_strides_0, weight = squeeze_16, x = var_3398)[name = string("op_3406")];
            tensor<int32, [3]> var_3407 = const()[name = string("op_3407"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_101 = transpose(perm = var_3407, x = var_3406)[name = string("transpose_101")];
            tensor<fp16, [1, 1024, 1024]> x_571_cast_fp16 = add(x = hidden_states_33_cast_fp16, y = attn_out_101)[name = string("x_571_cast_fp16")];
            fp16 var_5_promoted_67_to_fp16 = const()[name = string("op_5_promoted_67_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_3413_cast_fp16 = pow(x = x_571_cast_fp16, y = var_5_promoted_67_to_fp16)[name = string("op_3413_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("var_135_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_135_keep_dims_0 = const()[name = string("var_135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_135_cast_fp16 = reduce_mean(axes = var_135_axes_0, keep_dims = var_135_keep_dims_0, x = var_3413_cast_fp16)[name = string("var_135_cast_fp16")];
            fp16 var_3416_to_fp16 = const()[name = string("op_3416_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3417_cast_fp16 = add(x = var_135_cast_fp16, y = var_3416_to_fp16)[name = string("op_3417_cast_fp16")];
            fp32 var_3418_epsilon_0 = const()[name = string("op_3418_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3418_cast_fp16 = rsqrt(epsilon = var_3418_epsilon_0, x = var_3417_cast_fp16)[name = string("op_3418_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_575_cast_fp16 = mul(x = x_571_cast_fp16, y = var_3418_cast_fp16)[name = string("x_575_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_16_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_16_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1162708352)))];
            tensor<fp16, [1, 1024, 1024]> var_3421_cast_fp16 = mul(x = x_575_cast_fp16, y = encoder_layers_16_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_3421_cast_fp16")];
            tensor<int32, [3]> var_3426 = const()[name = string("op_3426"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_165_axes_0 = const()[name = string("input_165_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3427 = transpose(perm = var_3426, x = var_3421_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 1024, 1, 1024]> input_165 = expand_dims(axes = input_165_axes_0, x = var_3427)[name = string("input_165")];
            string input_167_pad_type_0 = const()[name = string("input_167_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_167_strides_0 = const()[name = string("input_167_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_167_pad_0 = const()[name = string("input_167_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_167_dilations_0 = const()[name = string("input_167_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_167_groups_0 = const()[name = string("input_167_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_167 = conv(dilations = input_167_dilations_0, groups = input_167_groups_0, pad = input_167_pad_0, pad_type = input_167_pad_type_0, strides = input_167_strides_0, weight = encoder_layers_16_mlp_gate_proj_weight, x = input_165)[name = string("input_167")];
            string up_33_pad_type_0 = const()[name = string("up_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_33_strides_0 = const()[name = string("up_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_33_pad_0 = const()[name = string("up_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_33_dilations_0 = const()[name = string("up_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_33_groups_0 = const()[name = string("up_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_33 = conv(dilations = up_33_dilations_0, groups = up_33_groups_0, pad = up_33_pad_0, pad_type = up_33_pad_type_0, strides = up_33_strides_0, weight = encoder_layers_16_mlp_up_proj_weight, x = input_165)[name = string("up_33")];
            tensor<fp16, [1, 3072, 1, 1024]> var_3441 = silu(x = input_167)[name = string("op_3441")];
            tensor<fp16, [1, 3072, 1, 1024]> input_169 = mul(x = var_3441, y = up_33)[name = string("input_169")];
            string var_3448_pad_type_0 = const()[name = string("op_3448_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3448_strides_0 = const()[name = string("op_3448_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3448_pad_0 = const()[name = string("op_3448_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3448_dilations_0 = const()[name = string("op_3448_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3448_groups_0 = const()[name = string("op_3448_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3448 = conv(dilations = var_3448_dilations_0, groups = var_3448_groups_0, pad = var_3448_pad_0, pad_type = var_3448_pad_type_0, strides = var_3448_strides_0, weight = encoder_layers_16_mlp_down_proj_weight, x = input_169)[name = string("op_3448")];
            tensor<int32, [1]> var_3449_axes_0 = const()[name = string("op_3449_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3449 = squeeze(axes = var_3449_axes_0, x = var_3448)[name = string("op_3449")];
            tensor<int32, [3]> var_3450 = const()[name = string("op_3450"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_33 = transpose(perm = var_3450, x = var_3449)[name = string("transpose_99")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_35_cast_fp16 = add(x = x_571_cast_fp16, y = mlp_out_33)[name = string("hidden_states_35_cast_fp16")];
            fp16 var_5_promoted_68_to_fp16 = const()[name = string("op_5_promoted_68_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_3477_cast_fp16 = pow(x = hidden_states_35_cast_fp16, y = var_5_promoted_68_to_fp16)[name = string("op_3477_cast_fp16")];
            tensor<int32, [1]> var_137_axes_0 = const()[name = string("var_137_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_137_keep_dims_0 = const()[name = string("var_137_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_137_cast_fp16 = reduce_mean(axes = var_137_axes_0, keep_dims = var_137_keep_dims_0, x = var_3477_cast_fp16)[name = string("var_137_cast_fp16")];
            fp16 var_3480_to_fp16 = const()[name = string("op_3480_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3481_cast_fp16 = add(x = var_137_cast_fp16, y = var_3480_to_fp16)[name = string("op_3481_cast_fp16")];
            fp32 var_3482_epsilon_0 = const()[name = string("op_3482_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3482_cast_fp16 = rsqrt(epsilon = var_3482_epsilon_0, x = var_3481_cast_fp16)[name = string("op_3482_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_581_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = var_3482_cast_fp16)[name = string("x_581_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_17_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_17_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1162710464)))];
            tensor<fp16, [1, 1024, 1024]> var_3485_cast_fp16 = mul(x = x_581_cast_fp16, y = encoder_layers_17_input_layernorm_weight_promoted_to_fp16)[name = string("op_3485_cast_fp16")];
            tensor<int32, [3]> var_3490 = const()[name = string("op_3490"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_171_axes_0 = const()[name = string("input_171_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3491 = transpose(perm = var_3490, x = var_3485_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 1024, 1, 1024]> input_171 = expand_dims(axes = input_171_axes_0, x = var_3491)[name = string("input_171")];
            string var_3498_pad_type_0 = const()[name = string("op_3498_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3498_strides_0 = const()[name = string("op_3498_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3498_pad_0 = const()[name = string("op_3498_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3498_dilations_0 = const()[name = string("op_3498_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3498_groups_0 = const()[name = string("op_3498_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_3498 = conv(dilations = var_3498_dilations_0, groups = var_3498_groups_0, pad = var_3498_pad_0, pad_type = var_3498_pad_type_0, strides = var_3498_strides_0, weight = encoder_layers_17_self_attn_q_proj_weight, x = input_171)[name = string("op_3498")];
            tensor<int32, [4]> var_3499 = const()[name = string("op_3499"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_3500 = reshape(shape = var_3499, x = var_3498)[name = string("op_3500")];
            tensor<int32, [4]> var_3501 = const()[name = string("op_3501"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3508_pad_type_0 = const()[name = string("op_3508_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3508_strides_0 = const()[name = string("op_3508_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3508_pad_0 = const()[name = string("op_3508_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3508_dilations_0 = const()[name = string("op_3508_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3508_groups_0 = const()[name = string("op_3508_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3508 = conv(dilations = var_3508_dilations_0, groups = var_3508_groups_0, pad = var_3508_pad_0, pad_type = var_3508_pad_type_0, strides = var_3508_strides_0, weight = encoder_layers_17_self_attn_k_proj_weight, x = input_171)[name = string("op_3508")];
            tensor<int32, [4]> var_3509 = const()[name = string("op_3509"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_3510 = reshape(shape = var_3509, x = var_3508)[name = string("op_3510")];
            tensor<int32, [4]> var_3511 = const()[name = string("op_3511"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3518_pad_type_0 = const()[name = string("op_3518_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3518_strides_0 = const()[name = string("op_3518_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3518_pad_0 = const()[name = string("op_3518_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3518_dilations_0 = const()[name = string("op_3518_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3518_groups_0 = const()[name = string("op_3518_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3518 = conv(dilations = var_3518_dilations_0, groups = var_3518_groups_0, pad = var_3518_pad_0, pad_type = var_3518_pad_type_0, strides = var_3518_strides_0, weight = encoder_layers_17_self_attn_v_proj_weight, x = input_171)[name = string("op_3518")];
            tensor<int32, [4]> var_3519 = const()[name = string("op_3519"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_3520 = reshape(shape = var_3519, x = var_3518)[name = string("op_3520")];
            tensor<int32, [4]> var_3521 = const()[name = string("op_3521"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_69_to_fp16 = const()[name = string("op_5_promoted_69_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_103 = transpose(perm = var_3501, x = var_3500)[name = string("transpose_97")];
            tensor<fp16, [1, 16, 1024, 128]> var_3527_cast_fp16 = pow(x = q_103, y = var_5_promoted_69_to_fp16)[name = string("op_3527_cast_fp16")];
            tensor<int32, [1]> var_139_axes_0 = const()[name = string("var_139_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_139_keep_dims_0 = const()[name = string("var_139_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_139_cast_fp16 = reduce_mean(axes = var_139_axes_0, keep_dims = var_139_keep_dims_0, x = var_3527_cast_fp16)[name = string("var_139_cast_fp16")];
            fp16 var_3530_to_fp16 = const()[name = string("op_3530_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_3531_cast_fp16 = add(x = var_139_cast_fp16, y = var_3530_to_fp16)[name = string("op_3531_cast_fp16")];
            fp32 var_3532_epsilon_0 = const()[name = string("op_3532_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_3532_cast_fp16 = rsqrt(epsilon = var_3532_epsilon_0, x = var_3531_cast_fp16)[name = string("op_3532_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_589_cast_fp16 = mul(x = q_103, y = var_3532_cast_fp16)[name = string("x_589_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_105 = mul(x = x_589_cast_fp16, y = encoder_layers_17_self_attn_q_norm_weight)[name = string("q_105")];
            fp16 var_5_promoted_70_to_fp16 = const()[name = string("op_5_promoted_70_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_103 = transpose(perm = var_3511, x = var_3510)[name = string("transpose_96")];
            tensor<fp16, [1, 8, 1024, 128]> var_3540_cast_fp16 = pow(x = k_103, y = var_5_promoted_70_to_fp16)[name = string("op_3540_cast_fp16")];
            tensor<int32, [1]> var_141_axes_0 = const()[name = string("var_141_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_141_keep_dims_0 = const()[name = string("var_141_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_141_cast_fp16 = reduce_mean(axes = var_141_axes_0, keep_dims = var_141_keep_dims_0, x = var_3540_cast_fp16)[name = string("var_141_cast_fp16")];
            fp16 var_3543_to_fp16 = const()[name = string("op_3543_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_3544_cast_fp16 = add(x = var_141_cast_fp16, y = var_3543_to_fp16)[name = string("op_3544_cast_fp16")];
            fp32 var_3545_epsilon_0 = const()[name = string("op_3545_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_3545_cast_fp16 = rsqrt(epsilon = var_3545_epsilon_0, x = var_3544_cast_fp16)[name = string("op_3545_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_595_cast_fp16 = mul(x = k_103, y = var_3545_cast_fp16)[name = string("x_595_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_105 = mul(x = x_595_cast_fp16, y = encoder_layers_17_self_attn_k_norm_weight)[name = string("k_105")];
            tensor<fp16, [1, 16, 1024, 128]> var_3549 = mul(x = q_105, y = cos)[name = string("op_3549")];
            tensor<int32, [2]> var_3550_split_sizes_0 = const()[name = string("op_3550_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3550_axis_0 = const()[name = string("op_3550_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_3550_0, tensor<fp16, [1, 16, 1024, 64]> var_3550_1 = split(axis = var_3550_axis_0, split_sizes = var_3550_split_sizes_0, x = q_105)[name = string("op_3550")];
            fp16 const_54_promoted = const()[name = string("const_54_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_3552 = mul(x = var_3550_1, y = const_54_promoted)[name = string("op_3552")];
            bool var_3554_interleave_0 = const()[name = string("op_3554_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_3554 = concat(axis = var_17, interleave = var_3554_interleave_0, values = (var_3552, var_3550_0))[name = string("op_3554")];
            tensor<fp16, [1, 16, 1024, 128]> var_3555 = mul(x = var_3554, y = sin)[name = string("op_3555")];
            tensor<fp16, [1, 16, 1024, 128]> query_35 = add(x = var_3549, y = var_3555)[name = string("query_35")];
            tensor<fp16, [1, 8, 1024, 128]> var_3557 = mul(x = k_105, y = cos)[name = string("op_3557")];
            tensor<int32, [2]> var_3558_split_sizes_0 = const()[name = string("op_3558_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3558_axis_0 = const()[name = string("op_3558_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_3558_0, tensor<fp16, [1, 8, 1024, 64]> var_3558_1 = split(axis = var_3558_axis_0, split_sizes = var_3558_split_sizes_0, x = k_105)[name = string("op_3558")];
            fp16 const_55_promoted = const()[name = string("const_55_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_3560 = mul(x = var_3558_1, y = const_55_promoted)[name = string("op_3560")];
            bool var_3562_interleave_0 = const()[name = string("op_3562_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_3562 = concat(axis = var_17, interleave = var_3562_interleave_0, values = (var_3560, var_3558_0))[name = string("op_3562")];
            tensor<fp16, [1, 8, 1024, 128]> var_3563 = mul(x = var_3562, y = sin)[name = string("op_3563")];
            tensor<fp16, [1, 8, 1024, 128]> x_597 = add(x = var_3557, y = var_3563)[name = string("x_597")];
            tensor<int32, [1]> var_3565_axes_0 = const()[name = string("op_3565_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3565 = expand_dims(axes = var_3565_axes_0, x = x_597)[name = string("op_3565")];
            tensor<int32, [5]> x_599_reps_0 = const()[name = string("x_599_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_599 = tile(reps = x_599_reps_0, x = var_3565)[name = string("x_599")];
            tensor<int32, [4]> var_3568 = const()[name = string("op_3568"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_35 = reshape(shape = var_3568, x = x_599)[name = string("key_35")];
            tensor<int32, [1]> var_3570_axes_0 = const()[name = string("op_3570_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_601 = transpose(perm = var_3521, x = var_3520)[name = string("transpose_95")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3570 = expand_dims(axes = var_3570_axes_0, x = x_601)[name = string("op_3570")];
            tensor<int32, [5]> x_603_reps_0 = const()[name = string("x_603_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_603 = tile(reps = x_603_reps_0, x = var_3570)[name = string("x_603")];
            tensor<int32, [4]> var_3573 = const()[name = string("op_3573"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_35 = reshape(shape = var_3573, x = x_603)[name = string("value_35")];
            bool var_3578_transpose_x_1 = const()[name = string("op_3578_transpose_x_1"), val = bool(false)];
            bool var_3578_transpose_y_1 = const()[name = string("op_3578_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_3578_cast_fp16 = matmul(transpose_x = var_3578_transpose_x_1, transpose_y = var_3578_transpose_y_1, x = query_35, y = key_35)[name = string("op_3578_cast_fp16")];
            fp16 var_3579_to_fp16 = const()[name = string("op_3579_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_103_cast_fp16 = mul(x = var_3578_cast_fp16, y = var_3579_to_fp16)[name = string("attn_weights_103_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_105_cast_fp16 = add(x = attn_weights_103_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_105_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_3583_cast_fp16 = softmax(axis = var_17, x = attn_weights_105_cast_fp16)[name = string("op_3583_cast_fp16")];
            bool var_3587_transpose_x_0 = const()[name = string("op_3587_transpose_x_0"), val = bool(false)];
            bool var_3587_transpose_y_0 = const()[name = string("op_3587_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_3587_cast_fp16 = matmul(transpose_x = var_3587_transpose_x_0, transpose_y = var_3587_transpose_y_0, x = var_3583_cast_fp16, y = value_35)[name = string("op_3587_cast_fp16")];
            tensor<int32, [4]> var_3589 = const()[name = string("op_3589"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3592 = const()[name = string("op_3592"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_3590 = transpose(perm = var_3589, x = var_3587_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 1024, 2048]> attn_out_105 = reshape(shape = var_3592, x = var_3590)[name = string("attn_out_105")];
            tensor<int32, [3]> var_3594 = const()[name = string("op_3594"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_17 = const()[name = string("squeeze_17"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1162712576)))];
            string var_3603_pad_type_0 = const()[name = string("op_3603_pad_type_0"), val = string("valid")];
            int32 var_3603_groups_0 = const()[name = string("op_3603_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3603_strides_0 = const()[name = string("op_3603_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3603_pad_0 = const()[name = string("op_3603_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3603_dilations_0 = const()[name = string("op_3603_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_3595 = transpose(perm = var_3594, x = attn_out_105)[name = string("transpose_93")];
            tensor<fp16, [1, 1024, 1024]> var_3603 = conv(dilations = var_3603_dilations_0, groups = var_3603_groups_0, pad = var_3603_pad_0, pad_type = var_3603_pad_type_0, strides = var_3603_strides_0, weight = squeeze_17, x = var_3595)[name = string("op_3603")];
            tensor<int32, [3]> var_3604 = const()[name = string("op_3604"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_107 = transpose(perm = var_3604, x = var_3603)[name = string("transpose_92")];
            tensor<fp16, [1, 1024, 1024]> x_605_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = attn_out_107)[name = string("x_605_cast_fp16")];
            fp16 var_5_promoted_71_to_fp16 = const()[name = string("op_5_promoted_71_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_3610_cast_fp16 = pow(x = x_605_cast_fp16, y = var_5_promoted_71_to_fp16)[name = string("op_3610_cast_fp16")];
            tensor<int32, [1]> var_143_axes_0 = const()[name = string("var_143_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_143_keep_dims_0 = const()[name = string("var_143_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_143_cast_fp16 = reduce_mean(axes = var_143_axes_0, keep_dims = var_143_keep_dims_0, x = var_3610_cast_fp16)[name = string("var_143_cast_fp16")];
            fp16 var_3613_to_fp16 = const()[name = string("op_3613_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3614_cast_fp16 = add(x = var_143_cast_fp16, y = var_3613_to_fp16)[name = string("op_3614_cast_fp16")];
            fp32 var_3615_epsilon_0 = const()[name = string("op_3615_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3615_cast_fp16 = rsqrt(epsilon = var_3615_epsilon_0, x = var_3614_cast_fp16)[name = string("op_3615_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_609_cast_fp16 = mul(x = x_605_cast_fp16, y = var_3615_cast_fp16)[name = string("x_609_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_17_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_17_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1166906944)))];
            tensor<fp16, [1, 1024, 1024]> var_3618_cast_fp16 = mul(x = x_609_cast_fp16, y = encoder_layers_17_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_3618_cast_fp16")];
            tensor<int32, [3]> var_3623 = const()[name = string("op_3623"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_175_axes_0 = const()[name = string("input_175_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3624 = transpose(perm = var_3623, x = var_3618_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 1024, 1, 1024]> input_175 = expand_dims(axes = input_175_axes_0, x = var_3624)[name = string("input_175")];
            string input_177_pad_type_0 = const()[name = string("input_177_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_177_strides_0 = const()[name = string("input_177_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_177_pad_0 = const()[name = string("input_177_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_177_dilations_0 = const()[name = string("input_177_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_177_groups_0 = const()[name = string("input_177_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_177 = conv(dilations = input_177_dilations_0, groups = input_177_groups_0, pad = input_177_pad_0, pad_type = input_177_pad_type_0, strides = input_177_strides_0, weight = encoder_layers_17_mlp_gate_proj_weight, x = input_175)[name = string("input_177")];
            string up_35_pad_type_0 = const()[name = string("up_35_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_35_strides_0 = const()[name = string("up_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_35_pad_0 = const()[name = string("up_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_35_dilations_0 = const()[name = string("up_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_35_groups_0 = const()[name = string("up_35_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_35 = conv(dilations = up_35_dilations_0, groups = up_35_groups_0, pad = up_35_pad_0, pad_type = up_35_pad_type_0, strides = up_35_strides_0, weight = encoder_layers_17_mlp_up_proj_weight, x = input_175)[name = string("up_35")];
            tensor<fp16, [1, 3072, 1, 1024]> var_3638 = silu(x = input_177)[name = string("op_3638")];
            tensor<fp16, [1, 3072, 1, 1024]> input_179 = mul(x = var_3638, y = up_35)[name = string("input_179")];
            string var_3645_pad_type_0 = const()[name = string("op_3645_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3645_strides_0 = const()[name = string("op_3645_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3645_pad_0 = const()[name = string("op_3645_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3645_dilations_0 = const()[name = string("op_3645_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3645_groups_0 = const()[name = string("op_3645_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3645 = conv(dilations = var_3645_dilations_0, groups = var_3645_groups_0, pad = var_3645_pad_0, pad_type = var_3645_pad_type_0, strides = var_3645_strides_0, weight = encoder_layers_17_mlp_down_proj_weight, x = input_179)[name = string("op_3645")];
            tensor<int32, [1]> var_3646_axes_0 = const()[name = string("op_3646_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3646 = squeeze(axes = var_3646_axes_0, x = var_3645)[name = string("op_3646")];
            tensor<int32, [3]> var_3647 = const()[name = string("op_3647"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_35 = transpose(perm = var_3647, x = var_3646)[name = string("transpose_90")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_37_cast_fp16 = add(x = x_605_cast_fp16, y = mlp_out_35)[name = string("hidden_states_37_cast_fp16")];
            fp16 var_5_promoted_72_to_fp16 = const()[name = string("op_5_promoted_72_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_3674_cast_fp16 = pow(x = hidden_states_37_cast_fp16, y = var_5_promoted_72_to_fp16)[name = string("op_3674_cast_fp16")];
            tensor<int32, [1]> var_145_axes_0 = const()[name = string("var_145_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_145_keep_dims_0 = const()[name = string("var_145_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_145_cast_fp16 = reduce_mean(axes = var_145_axes_0, keep_dims = var_145_keep_dims_0, x = var_3674_cast_fp16)[name = string("var_145_cast_fp16")];
            fp16 var_3677_to_fp16 = const()[name = string("op_3677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3678_cast_fp16 = add(x = var_145_cast_fp16, y = var_3677_to_fp16)[name = string("op_3678_cast_fp16")];
            fp32 var_3679_epsilon_0 = const()[name = string("op_3679_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3679_cast_fp16 = rsqrt(epsilon = var_3679_epsilon_0, x = var_3678_cast_fp16)[name = string("op_3679_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_615_cast_fp16 = mul(x = hidden_states_37_cast_fp16, y = var_3679_cast_fp16)[name = string("x_615_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_18_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_18_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1166909056)))];
            tensor<fp16, [1, 1024, 1024]> var_3682_cast_fp16 = mul(x = x_615_cast_fp16, y = encoder_layers_18_input_layernorm_weight_promoted_to_fp16)[name = string("op_3682_cast_fp16")];
            tensor<int32, [3]> var_3687 = const()[name = string("op_3687"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3688 = transpose(perm = var_3687, x = var_3682_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 1024, 1, 1024]> input_181 = expand_dims(axes = input_181_axes_0, x = var_3688)[name = string("input_181")];
            string var_3695_pad_type_0 = const()[name = string("op_3695_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3695_strides_0 = const()[name = string("op_3695_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3695_pad_0 = const()[name = string("op_3695_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3695_dilations_0 = const()[name = string("op_3695_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3695_groups_0 = const()[name = string("op_3695_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_3695 = conv(dilations = var_3695_dilations_0, groups = var_3695_groups_0, pad = var_3695_pad_0, pad_type = var_3695_pad_type_0, strides = var_3695_strides_0, weight = encoder_layers_18_self_attn_q_proj_weight, x = input_181)[name = string("op_3695")];
            tensor<int32, [4]> var_3696 = const()[name = string("op_3696"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_3697 = reshape(shape = var_3696, x = var_3695)[name = string("op_3697")];
            tensor<int32, [4]> var_3698 = const()[name = string("op_3698"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3705_pad_type_0 = const()[name = string("op_3705_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3705_strides_0 = const()[name = string("op_3705_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3705_pad_0 = const()[name = string("op_3705_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3705_dilations_0 = const()[name = string("op_3705_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3705_groups_0 = const()[name = string("op_3705_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3705 = conv(dilations = var_3705_dilations_0, groups = var_3705_groups_0, pad = var_3705_pad_0, pad_type = var_3705_pad_type_0, strides = var_3705_strides_0, weight = encoder_layers_18_self_attn_k_proj_weight, x = input_181)[name = string("op_3705")];
            tensor<int32, [4]> var_3706 = const()[name = string("op_3706"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_3707 = reshape(shape = var_3706, x = var_3705)[name = string("op_3707")];
            tensor<int32, [4]> var_3708 = const()[name = string("op_3708"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3715_pad_type_0 = const()[name = string("op_3715_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3715_strides_0 = const()[name = string("op_3715_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3715_pad_0 = const()[name = string("op_3715_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3715_dilations_0 = const()[name = string("op_3715_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3715_groups_0 = const()[name = string("op_3715_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3715 = conv(dilations = var_3715_dilations_0, groups = var_3715_groups_0, pad = var_3715_pad_0, pad_type = var_3715_pad_type_0, strides = var_3715_strides_0, weight = encoder_layers_18_self_attn_v_proj_weight, x = input_181)[name = string("op_3715")];
            tensor<int32, [4]> var_3716 = const()[name = string("op_3716"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_3717 = reshape(shape = var_3716, x = var_3715)[name = string("op_3717")];
            tensor<int32, [4]> var_3718 = const()[name = string("op_3718"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_73_to_fp16 = const()[name = string("op_5_promoted_73_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_109 = transpose(perm = var_3698, x = var_3697)[name = string("transpose_88")];
            tensor<fp16, [1, 16, 1024, 128]> var_3724_cast_fp16 = pow(x = q_109, y = var_5_promoted_73_to_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_147_axes_0 = const()[name = string("var_147_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_147_keep_dims_0 = const()[name = string("var_147_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_147_cast_fp16 = reduce_mean(axes = var_147_axes_0, keep_dims = var_147_keep_dims_0, x = var_3724_cast_fp16)[name = string("var_147_cast_fp16")];
            fp16 var_3727_to_fp16 = const()[name = string("op_3727_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_3728_cast_fp16 = add(x = var_147_cast_fp16, y = var_3727_to_fp16)[name = string("op_3728_cast_fp16")];
            fp32 var_3729_epsilon_0 = const()[name = string("op_3729_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_3729_cast_fp16 = rsqrt(epsilon = var_3729_epsilon_0, x = var_3728_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_623_cast_fp16 = mul(x = q_109, y = var_3729_cast_fp16)[name = string("x_623_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_111 = mul(x = x_623_cast_fp16, y = encoder_layers_18_self_attn_q_norm_weight)[name = string("q_111")];
            fp16 var_5_promoted_74_to_fp16 = const()[name = string("op_5_promoted_74_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_109 = transpose(perm = var_3708, x = var_3707)[name = string("transpose_87")];
            tensor<fp16, [1, 8, 1024, 128]> var_3737_cast_fp16 = pow(x = k_109, y = var_5_promoted_74_to_fp16)[name = string("op_3737_cast_fp16")];
            tensor<int32, [1]> var_149_axes_0 = const()[name = string("var_149_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_149_keep_dims_0 = const()[name = string("var_149_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_149_cast_fp16 = reduce_mean(axes = var_149_axes_0, keep_dims = var_149_keep_dims_0, x = var_3737_cast_fp16)[name = string("var_149_cast_fp16")];
            fp16 var_3740_to_fp16 = const()[name = string("op_3740_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_3741_cast_fp16 = add(x = var_149_cast_fp16, y = var_3740_to_fp16)[name = string("op_3741_cast_fp16")];
            fp32 var_3742_epsilon_0 = const()[name = string("op_3742_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_3742_cast_fp16 = rsqrt(epsilon = var_3742_epsilon_0, x = var_3741_cast_fp16)[name = string("op_3742_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_629_cast_fp16 = mul(x = k_109, y = var_3742_cast_fp16)[name = string("x_629_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_111 = mul(x = x_629_cast_fp16, y = encoder_layers_18_self_attn_k_norm_weight)[name = string("k_111")];
            tensor<fp16, [1, 16, 1024, 128]> var_3746 = mul(x = q_111, y = cos)[name = string("op_3746")];
            tensor<int32, [2]> var_3747_split_sizes_0 = const()[name = string("op_3747_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3747_axis_0 = const()[name = string("op_3747_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_3747_0, tensor<fp16, [1, 16, 1024, 64]> var_3747_1 = split(axis = var_3747_axis_0, split_sizes = var_3747_split_sizes_0, x = q_111)[name = string("op_3747")];
            fp16 const_57_promoted = const()[name = string("const_57_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_3749 = mul(x = var_3747_1, y = const_57_promoted)[name = string("op_3749")];
            bool var_3751_interleave_0 = const()[name = string("op_3751_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_3751 = concat(axis = var_17, interleave = var_3751_interleave_0, values = (var_3749, var_3747_0))[name = string("op_3751")];
            tensor<fp16, [1, 16, 1024, 128]> var_3752 = mul(x = var_3751, y = sin)[name = string("op_3752")];
            tensor<fp16, [1, 16, 1024, 128]> query_37 = add(x = var_3746, y = var_3752)[name = string("query_37")];
            tensor<fp16, [1, 8, 1024, 128]> var_3754 = mul(x = k_111, y = cos)[name = string("op_3754")];
            tensor<int32, [2]> var_3755_split_sizes_0 = const()[name = string("op_3755_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3755_axis_0 = const()[name = string("op_3755_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_3755_0, tensor<fp16, [1, 8, 1024, 64]> var_3755_1 = split(axis = var_3755_axis_0, split_sizes = var_3755_split_sizes_0, x = k_111)[name = string("op_3755")];
            fp16 const_58_promoted = const()[name = string("const_58_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_3757 = mul(x = var_3755_1, y = const_58_promoted)[name = string("op_3757")];
            bool var_3759_interleave_0 = const()[name = string("op_3759_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_3759 = concat(axis = var_17, interleave = var_3759_interleave_0, values = (var_3757, var_3755_0))[name = string("op_3759")];
            tensor<fp16, [1, 8, 1024, 128]> var_3760 = mul(x = var_3759, y = sin)[name = string("op_3760")];
            tensor<fp16, [1, 8, 1024, 128]> x_631 = add(x = var_3754, y = var_3760)[name = string("x_631")];
            tensor<int32, [1]> var_3762_axes_0 = const()[name = string("op_3762_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3762 = expand_dims(axes = var_3762_axes_0, x = x_631)[name = string("op_3762")];
            tensor<int32, [5]> x_633_reps_0 = const()[name = string("x_633_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_633 = tile(reps = x_633_reps_0, x = var_3762)[name = string("x_633")];
            tensor<int32, [4]> var_3765 = const()[name = string("op_3765"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_37 = reshape(shape = var_3765, x = x_633)[name = string("key_37")];
            tensor<int32, [1]> var_3767_axes_0 = const()[name = string("op_3767_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_635 = transpose(perm = var_3718, x = var_3717)[name = string("transpose_86")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3767 = expand_dims(axes = var_3767_axes_0, x = x_635)[name = string("op_3767")];
            tensor<int32, [5]> x_637_reps_0 = const()[name = string("x_637_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_637 = tile(reps = x_637_reps_0, x = var_3767)[name = string("x_637")];
            tensor<int32, [4]> var_3770 = const()[name = string("op_3770"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_37 = reshape(shape = var_3770, x = x_637)[name = string("value_37")];
            bool var_3775_transpose_x_1 = const()[name = string("op_3775_transpose_x_1"), val = bool(false)];
            bool var_3775_transpose_y_1 = const()[name = string("op_3775_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_3775_cast_fp16 = matmul(transpose_x = var_3775_transpose_x_1, transpose_y = var_3775_transpose_y_1, x = query_37, y = key_37)[name = string("op_3775_cast_fp16")];
            fp16 var_3776_to_fp16 = const()[name = string("op_3776_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_109_cast_fp16 = mul(x = var_3775_cast_fp16, y = var_3776_to_fp16)[name = string("attn_weights_109_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_111_cast_fp16 = add(x = attn_weights_109_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_3780_cast_fp16 = softmax(axis = var_17, x = attn_weights_111_cast_fp16)[name = string("op_3780_cast_fp16")];
            bool var_3784_transpose_x_0 = const()[name = string("op_3784_transpose_x_0"), val = bool(false)];
            bool var_3784_transpose_y_0 = const()[name = string("op_3784_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_3784_cast_fp16 = matmul(transpose_x = var_3784_transpose_x_0, transpose_y = var_3784_transpose_y_0, x = var_3780_cast_fp16, y = value_37)[name = string("op_3784_cast_fp16")];
            tensor<int32, [4]> var_3786 = const()[name = string("op_3786"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3789 = const()[name = string("op_3789"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_3787 = transpose(perm = var_3786, x = var_3784_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 1024, 2048]> attn_out_111 = reshape(shape = var_3789, x = var_3787)[name = string("attn_out_111")];
            tensor<int32, [3]> var_3791 = const()[name = string("op_3791"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_18 = const()[name = string("squeeze_18"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1166911168)))];
            string var_3800_pad_type_0 = const()[name = string("op_3800_pad_type_0"), val = string("valid")];
            int32 var_3800_groups_0 = const()[name = string("op_3800_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3800_strides_0 = const()[name = string("op_3800_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3800_pad_0 = const()[name = string("op_3800_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3800_dilations_0 = const()[name = string("op_3800_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_3792 = transpose(perm = var_3791, x = attn_out_111)[name = string("transpose_84")];
            tensor<fp16, [1, 1024, 1024]> var_3800 = conv(dilations = var_3800_dilations_0, groups = var_3800_groups_0, pad = var_3800_pad_0, pad_type = var_3800_pad_type_0, strides = var_3800_strides_0, weight = squeeze_18, x = var_3792)[name = string("op_3800")];
            tensor<int32, [3]> var_3801 = const()[name = string("op_3801"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_113 = transpose(perm = var_3801, x = var_3800)[name = string("transpose_83")];
            tensor<fp16, [1, 1024, 1024]> x_639_cast_fp16 = add(x = hidden_states_37_cast_fp16, y = attn_out_113)[name = string("x_639_cast_fp16")];
            fp16 var_5_promoted_75_to_fp16 = const()[name = string("op_5_promoted_75_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_3807_cast_fp16 = pow(x = x_639_cast_fp16, y = var_5_promoted_75_to_fp16)[name = string("op_3807_cast_fp16")];
            tensor<int32, [1]> var_151_axes_0 = const()[name = string("var_151_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_151_keep_dims_0 = const()[name = string("var_151_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_151_cast_fp16 = reduce_mean(axes = var_151_axes_0, keep_dims = var_151_keep_dims_0, x = var_3807_cast_fp16)[name = string("var_151_cast_fp16")];
            fp16 var_3810_to_fp16 = const()[name = string("op_3810_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3811_cast_fp16 = add(x = var_151_cast_fp16, y = var_3810_to_fp16)[name = string("op_3811_cast_fp16")];
            fp32 var_3812_epsilon_0 = const()[name = string("op_3812_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3812_cast_fp16 = rsqrt(epsilon = var_3812_epsilon_0, x = var_3811_cast_fp16)[name = string("op_3812_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_643_cast_fp16 = mul(x = x_639_cast_fp16, y = var_3812_cast_fp16)[name = string("x_643_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_18_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_18_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1171105536)))];
            tensor<fp16, [1, 1024, 1024]> var_3815_cast_fp16 = mul(x = x_643_cast_fp16, y = encoder_layers_18_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_3815_cast_fp16")];
            tensor<int32, [3]> var_3820 = const()[name = string("op_3820"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_185_axes_0 = const()[name = string("input_185_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3821 = transpose(perm = var_3820, x = var_3815_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 1024, 1, 1024]> input_185 = expand_dims(axes = input_185_axes_0, x = var_3821)[name = string("input_185")];
            string input_187_pad_type_0 = const()[name = string("input_187_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_187_strides_0 = const()[name = string("input_187_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_187_pad_0 = const()[name = string("input_187_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_187_dilations_0 = const()[name = string("input_187_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_187_groups_0 = const()[name = string("input_187_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_187 = conv(dilations = input_187_dilations_0, groups = input_187_groups_0, pad = input_187_pad_0, pad_type = input_187_pad_type_0, strides = input_187_strides_0, weight = encoder_layers_18_mlp_gate_proj_weight, x = input_185)[name = string("input_187")];
            string up_37_pad_type_0 = const()[name = string("up_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_37_strides_0 = const()[name = string("up_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_37_pad_0 = const()[name = string("up_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_37_dilations_0 = const()[name = string("up_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_37_groups_0 = const()[name = string("up_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_37 = conv(dilations = up_37_dilations_0, groups = up_37_groups_0, pad = up_37_pad_0, pad_type = up_37_pad_type_0, strides = up_37_strides_0, weight = encoder_layers_18_mlp_up_proj_weight, x = input_185)[name = string("up_37")];
            tensor<fp16, [1, 3072, 1, 1024]> var_3835 = silu(x = input_187)[name = string("op_3835")];
            tensor<fp16, [1, 3072, 1, 1024]> input_189 = mul(x = var_3835, y = up_37)[name = string("input_189")];
            string var_3842_pad_type_0 = const()[name = string("op_3842_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3842_strides_0 = const()[name = string("op_3842_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3842_pad_0 = const()[name = string("op_3842_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3842_dilations_0 = const()[name = string("op_3842_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3842_groups_0 = const()[name = string("op_3842_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3842 = conv(dilations = var_3842_dilations_0, groups = var_3842_groups_0, pad = var_3842_pad_0, pad_type = var_3842_pad_type_0, strides = var_3842_strides_0, weight = encoder_layers_18_mlp_down_proj_weight, x = input_189)[name = string("op_3842")];
            tensor<int32, [1]> var_3843_axes_0 = const()[name = string("op_3843_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3843 = squeeze(axes = var_3843_axes_0, x = var_3842)[name = string("op_3843")];
            tensor<int32, [3]> var_3844 = const()[name = string("op_3844"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_37 = transpose(perm = var_3844, x = var_3843)[name = string("transpose_81")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_39_cast_fp16 = add(x = x_639_cast_fp16, y = mlp_out_37)[name = string("hidden_states_39_cast_fp16")];
            fp16 var_5_promoted_76_to_fp16 = const()[name = string("op_5_promoted_76_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_3871_cast_fp16 = pow(x = hidden_states_39_cast_fp16, y = var_5_promoted_76_to_fp16)[name = string("op_3871_cast_fp16")];
            tensor<int32, [1]> var_153_axes_0 = const()[name = string("var_153_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_153_keep_dims_0 = const()[name = string("var_153_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_153_cast_fp16 = reduce_mean(axes = var_153_axes_0, keep_dims = var_153_keep_dims_0, x = var_3871_cast_fp16)[name = string("var_153_cast_fp16")];
            fp16 var_3874_to_fp16 = const()[name = string("op_3874_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3875_cast_fp16 = add(x = var_153_cast_fp16, y = var_3874_to_fp16)[name = string("op_3875_cast_fp16")];
            fp32 var_3876_epsilon_0 = const()[name = string("op_3876_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3876_cast_fp16 = rsqrt(epsilon = var_3876_epsilon_0, x = var_3875_cast_fp16)[name = string("op_3876_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_649_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = var_3876_cast_fp16)[name = string("x_649_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_19_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_19_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1171107648)))];
            tensor<fp16, [1, 1024, 1024]> var_3879_cast_fp16 = mul(x = x_649_cast_fp16, y = encoder_layers_19_input_layernorm_weight_promoted_to_fp16)[name = string("op_3879_cast_fp16")];
            tensor<int32, [3]> var_3884 = const()[name = string("op_3884"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_191_axes_0 = const()[name = string("input_191_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3885 = transpose(perm = var_3884, x = var_3879_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 1024, 1, 1024]> input_191 = expand_dims(axes = input_191_axes_0, x = var_3885)[name = string("input_191")];
            string var_3892_pad_type_0 = const()[name = string("op_3892_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3892_strides_0 = const()[name = string("op_3892_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3892_pad_0 = const()[name = string("op_3892_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3892_dilations_0 = const()[name = string("op_3892_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3892_groups_0 = const()[name = string("op_3892_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_3892 = conv(dilations = var_3892_dilations_0, groups = var_3892_groups_0, pad = var_3892_pad_0, pad_type = var_3892_pad_type_0, strides = var_3892_strides_0, weight = encoder_layers_19_self_attn_q_proj_weight, x = input_191)[name = string("op_3892")];
            tensor<int32, [4]> var_3893 = const()[name = string("op_3893"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_3894 = reshape(shape = var_3893, x = var_3892)[name = string("op_3894")];
            tensor<int32, [4]> var_3895 = const()[name = string("op_3895"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3902_pad_type_0 = const()[name = string("op_3902_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3902_strides_0 = const()[name = string("op_3902_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3902_pad_0 = const()[name = string("op_3902_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3902_dilations_0 = const()[name = string("op_3902_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3902_groups_0 = const()[name = string("op_3902_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3902 = conv(dilations = var_3902_dilations_0, groups = var_3902_groups_0, pad = var_3902_pad_0, pad_type = var_3902_pad_type_0, strides = var_3902_strides_0, weight = encoder_layers_19_self_attn_k_proj_weight, x = input_191)[name = string("op_3902")];
            tensor<int32, [4]> var_3903 = const()[name = string("op_3903"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_3904 = reshape(shape = var_3903, x = var_3902)[name = string("op_3904")];
            tensor<int32, [4]> var_3905 = const()[name = string("op_3905"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3912_pad_type_0 = const()[name = string("op_3912_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3912_strides_0 = const()[name = string("op_3912_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3912_pad_0 = const()[name = string("op_3912_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3912_dilations_0 = const()[name = string("op_3912_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3912_groups_0 = const()[name = string("op_3912_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_3912 = conv(dilations = var_3912_dilations_0, groups = var_3912_groups_0, pad = var_3912_pad_0, pad_type = var_3912_pad_type_0, strides = var_3912_strides_0, weight = encoder_layers_19_self_attn_v_proj_weight, x = input_191)[name = string("op_3912")];
            tensor<int32, [4]> var_3913 = const()[name = string("op_3913"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_3914 = reshape(shape = var_3913, x = var_3912)[name = string("op_3914")];
            tensor<int32, [4]> var_3915 = const()[name = string("op_3915"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_77_to_fp16 = const()[name = string("op_5_promoted_77_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_115 = transpose(perm = var_3895, x = var_3894)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 1024, 128]> var_3921_cast_fp16 = pow(x = q_115, y = var_5_promoted_77_to_fp16)[name = string("op_3921_cast_fp16")];
            tensor<int32, [1]> var_155_axes_0 = const()[name = string("var_155_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_155_keep_dims_0 = const()[name = string("var_155_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_155_cast_fp16 = reduce_mean(axes = var_155_axes_0, keep_dims = var_155_keep_dims_0, x = var_3921_cast_fp16)[name = string("var_155_cast_fp16")];
            fp16 var_3924_to_fp16 = const()[name = string("op_3924_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_3925_cast_fp16 = add(x = var_155_cast_fp16, y = var_3924_to_fp16)[name = string("op_3925_cast_fp16")];
            fp32 var_3926_epsilon_0 = const()[name = string("op_3926_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_3926_cast_fp16 = rsqrt(epsilon = var_3926_epsilon_0, x = var_3925_cast_fp16)[name = string("op_3926_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_657_cast_fp16 = mul(x = q_115, y = var_3926_cast_fp16)[name = string("x_657_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_117 = mul(x = x_657_cast_fp16, y = encoder_layers_19_self_attn_q_norm_weight)[name = string("q_117")];
            fp16 var_5_promoted_78_to_fp16 = const()[name = string("op_5_promoted_78_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_115 = transpose(perm = var_3905, x = var_3904)[name = string("transpose_78")];
            tensor<fp16, [1, 8, 1024, 128]> var_3934_cast_fp16 = pow(x = k_115, y = var_5_promoted_78_to_fp16)[name = string("op_3934_cast_fp16")];
            tensor<int32, [1]> var_157_axes_0 = const()[name = string("var_157_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_157_keep_dims_0 = const()[name = string("var_157_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_157_cast_fp16 = reduce_mean(axes = var_157_axes_0, keep_dims = var_157_keep_dims_0, x = var_3934_cast_fp16)[name = string("var_157_cast_fp16")];
            fp16 var_3937_to_fp16 = const()[name = string("op_3937_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_3938_cast_fp16 = add(x = var_157_cast_fp16, y = var_3937_to_fp16)[name = string("op_3938_cast_fp16")];
            fp32 var_3939_epsilon_0 = const()[name = string("op_3939_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_3939_cast_fp16 = rsqrt(epsilon = var_3939_epsilon_0, x = var_3938_cast_fp16)[name = string("op_3939_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_663_cast_fp16 = mul(x = k_115, y = var_3939_cast_fp16)[name = string("x_663_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_117 = mul(x = x_663_cast_fp16, y = encoder_layers_19_self_attn_k_norm_weight)[name = string("k_117")];
            tensor<fp16, [1, 16, 1024, 128]> var_3943 = mul(x = q_117, y = cos)[name = string("op_3943")];
            tensor<int32, [2]> var_3944_split_sizes_0 = const()[name = string("op_3944_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3944_axis_0 = const()[name = string("op_3944_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_3944_0, tensor<fp16, [1, 16, 1024, 64]> var_3944_1 = split(axis = var_3944_axis_0, split_sizes = var_3944_split_sizes_0, x = q_117)[name = string("op_3944")];
            fp16 const_60_promoted = const()[name = string("const_60_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_3946 = mul(x = var_3944_1, y = const_60_promoted)[name = string("op_3946")];
            bool var_3948_interleave_0 = const()[name = string("op_3948_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_3948 = concat(axis = var_17, interleave = var_3948_interleave_0, values = (var_3946, var_3944_0))[name = string("op_3948")];
            tensor<fp16, [1, 16, 1024, 128]> var_3949 = mul(x = var_3948, y = sin)[name = string("op_3949")];
            tensor<fp16, [1, 16, 1024, 128]> query_39 = add(x = var_3943, y = var_3949)[name = string("query_39")];
            tensor<fp16, [1, 8, 1024, 128]> var_3951 = mul(x = k_117, y = cos)[name = string("op_3951")];
            tensor<int32, [2]> var_3952_split_sizes_0 = const()[name = string("op_3952_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3952_axis_0 = const()[name = string("op_3952_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_3952_0, tensor<fp16, [1, 8, 1024, 64]> var_3952_1 = split(axis = var_3952_axis_0, split_sizes = var_3952_split_sizes_0, x = k_117)[name = string("op_3952")];
            fp16 const_61_promoted = const()[name = string("const_61_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_3954 = mul(x = var_3952_1, y = const_61_promoted)[name = string("op_3954")];
            bool var_3956_interleave_0 = const()[name = string("op_3956_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_3956 = concat(axis = var_17, interleave = var_3956_interleave_0, values = (var_3954, var_3952_0))[name = string("op_3956")];
            tensor<fp16, [1, 8, 1024, 128]> var_3957 = mul(x = var_3956, y = sin)[name = string("op_3957")];
            tensor<fp16, [1, 8, 1024, 128]> x_665 = add(x = var_3951, y = var_3957)[name = string("x_665")];
            tensor<int32, [1]> var_3959_axes_0 = const()[name = string("op_3959_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3959 = expand_dims(axes = var_3959_axes_0, x = x_665)[name = string("op_3959")];
            tensor<int32, [5]> x_667_reps_0 = const()[name = string("x_667_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_667 = tile(reps = x_667_reps_0, x = var_3959)[name = string("x_667")];
            tensor<int32, [4]> var_3962 = const()[name = string("op_3962"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_39 = reshape(shape = var_3962, x = x_667)[name = string("key_39")];
            tensor<int32, [1]> var_3964_axes_0 = const()[name = string("op_3964_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_669 = transpose(perm = var_3915, x = var_3914)[name = string("transpose_77")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3964 = expand_dims(axes = var_3964_axes_0, x = x_669)[name = string("op_3964")];
            tensor<int32, [5]> x_671_reps_0 = const()[name = string("x_671_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_671 = tile(reps = x_671_reps_0, x = var_3964)[name = string("x_671")];
            tensor<int32, [4]> var_3967 = const()[name = string("op_3967"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_39 = reshape(shape = var_3967, x = x_671)[name = string("value_39")];
            bool var_3972_transpose_x_1 = const()[name = string("op_3972_transpose_x_1"), val = bool(false)];
            bool var_3972_transpose_y_1 = const()[name = string("op_3972_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_3972_cast_fp16 = matmul(transpose_x = var_3972_transpose_x_1, transpose_y = var_3972_transpose_y_1, x = query_39, y = key_39)[name = string("op_3972_cast_fp16")];
            fp16 var_3973_to_fp16 = const()[name = string("op_3973_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_115_cast_fp16 = mul(x = var_3972_cast_fp16, y = var_3973_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_3977_cast_fp16 = softmax(axis = var_17, x = attn_weights_117_cast_fp16)[name = string("op_3977_cast_fp16")];
            bool var_3981_transpose_x_0 = const()[name = string("op_3981_transpose_x_0"), val = bool(false)];
            bool var_3981_transpose_y_0 = const()[name = string("op_3981_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_3981_cast_fp16 = matmul(transpose_x = var_3981_transpose_x_0, transpose_y = var_3981_transpose_y_0, x = var_3977_cast_fp16, y = value_39)[name = string("op_3981_cast_fp16")];
            tensor<int32, [4]> var_3983 = const()[name = string("op_3983"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3986 = const()[name = string("op_3986"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_3984 = transpose(perm = var_3983, x = var_3981_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 1024, 2048]> attn_out_117 = reshape(shape = var_3986, x = var_3984)[name = string("attn_out_117")];
            tensor<int32, [3]> var_3988 = const()[name = string("op_3988"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_19 = const()[name = string("squeeze_19"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1171109760)))];
            string var_3997_pad_type_0 = const()[name = string("op_3997_pad_type_0"), val = string("valid")];
            int32 var_3997_groups_0 = const()[name = string("op_3997_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3997_strides_0 = const()[name = string("op_3997_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3997_pad_0 = const()[name = string("op_3997_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3997_dilations_0 = const()[name = string("op_3997_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_3989 = transpose(perm = var_3988, x = attn_out_117)[name = string("transpose_75")];
            tensor<fp16, [1, 1024, 1024]> var_3997 = conv(dilations = var_3997_dilations_0, groups = var_3997_groups_0, pad = var_3997_pad_0, pad_type = var_3997_pad_type_0, strides = var_3997_strides_0, weight = squeeze_19, x = var_3989)[name = string("op_3997")];
            tensor<int32, [3]> var_3998 = const()[name = string("op_3998"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_119 = transpose(perm = var_3998, x = var_3997)[name = string("transpose_74")];
            tensor<fp16, [1, 1024, 1024]> x_673_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = attn_out_119)[name = string("x_673_cast_fp16")];
            fp16 var_5_promoted_79_to_fp16 = const()[name = string("op_5_promoted_79_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_4004_cast_fp16 = pow(x = x_673_cast_fp16, y = var_5_promoted_79_to_fp16)[name = string("op_4004_cast_fp16")];
            tensor<int32, [1]> var_159_axes_0 = const()[name = string("var_159_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_159_keep_dims_0 = const()[name = string("var_159_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_159_cast_fp16 = reduce_mean(axes = var_159_axes_0, keep_dims = var_159_keep_dims_0, x = var_4004_cast_fp16)[name = string("var_159_cast_fp16")];
            fp16 var_4007_to_fp16 = const()[name = string("op_4007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_4008_cast_fp16 = add(x = var_159_cast_fp16, y = var_4007_to_fp16)[name = string("op_4008_cast_fp16")];
            fp32 var_4009_epsilon_0 = const()[name = string("op_4009_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_4009_cast_fp16 = rsqrt(epsilon = var_4009_epsilon_0, x = var_4008_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_677_cast_fp16 = mul(x = x_673_cast_fp16, y = var_4009_cast_fp16)[name = string("x_677_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_19_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_19_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1175304128)))];
            tensor<fp16, [1, 1024, 1024]> var_4012_cast_fp16 = mul(x = x_677_cast_fp16, y = encoder_layers_19_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4012_cast_fp16")];
            tensor<int32, [3]> var_4017 = const()[name = string("op_4017"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_195_axes_0 = const()[name = string("input_195_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4018 = transpose(perm = var_4017, x = var_4012_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 1024, 1, 1024]> input_195 = expand_dims(axes = input_195_axes_0, x = var_4018)[name = string("input_195")];
            string input_197_pad_type_0 = const()[name = string("input_197_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_197_strides_0 = const()[name = string("input_197_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_197_pad_0 = const()[name = string("input_197_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_197_dilations_0 = const()[name = string("input_197_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_197_groups_0 = const()[name = string("input_197_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_197 = conv(dilations = input_197_dilations_0, groups = input_197_groups_0, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = input_197_strides_0, weight = encoder_layers_19_mlp_gate_proj_weight, x = input_195)[name = string("input_197")];
            string up_39_pad_type_0 = const()[name = string("up_39_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_39_strides_0 = const()[name = string("up_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_39_pad_0 = const()[name = string("up_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_39_dilations_0 = const()[name = string("up_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_39_groups_0 = const()[name = string("up_39_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_39 = conv(dilations = up_39_dilations_0, groups = up_39_groups_0, pad = up_39_pad_0, pad_type = up_39_pad_type_0, strides = up_39_strides_0, weight = encoder_layers_19_mlp_up_proj_weight, x = input_195)[name = string("up_39")];
            tensor<fp16, [1, 3072, 1, 1024]> var_4032 = silu(x = input_197)[name = string("op_4032")];
            tensor<fp16, [1, 3072, 1, 1024]> input_199 = mul(x = var_4032, y = up_39)[name = string("input_199")];
            string var_4039_pad_type_0 = const()[name = string("op_4039_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4039_strides_0 = const()[name = string("op_4039_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4039_pad_0 = const()[name = string("op_4039_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4039_dilations_0 = const()[name = string("op_4039_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4039_groups_0 = const()[name = string("op_4039_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4039 = conv(dilations = var_4039_dilations_0, groups = var_4039_groups_0, pad = var_4039_pad_0, pad_type = var_4039_pad_type_0, strides = var_4039_strides_0, weight = encoder_layers_19_mlp_down_proj_weight, x = input_199)[name = string("op_4039")];
            tensor<int32, [1]> var_4040_axes_0 = const()[name = string("op_4040_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4040 = squeeze(axes = var_4040_axes_0, x = var_4039)[name = string("op_4040")];
            tensor<int32, [3]> var_4041 = const()[name = string("op_4041"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_39 = transpose(perm = var_4041, x = var_4040)[name = string("transpose_72")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_41_cast_fp16 = add(x = x_673_cast_fp16, y = mlp_out_39)[name = string("hidden_states_41_cast_fp16")];
            fp16 var_5_promoted_80_to_fp16 = const()[name = string("op_5_promoted_80_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_4068_cast_fp16 = pow(x = hidden_states_41_cast_fp16, y = var_5_promoted_80_to_fp16)[name = string("op_4068_cast_fp16")];
            tensor<int32, [1]> var_161_axes_0 = const()[name = string("var_161_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_161_keep_dims_0 = const()[name = string("var_161_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_161_cast_fp16 = reduce_mean(axes = var_161_axes_0, keep_dims = var_161_keep_dims_0, x = var_4068_cast_fp16)[name = string("var_161_cast_fp16")];
            fp16 var_4071_to_fp16 = const()[name = string("op_4071_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_4072_cast_fp16 = add(x = var_161_cast_fp16, y = var_4071_to_fp16)[name = string("op_4072_cast_fp16")];
            fp32 var_4073_epsilon_0 = const()[name = string("op_4073_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_4073_cast_fp16 = rsqrt(epsilon = var_4073_epsilon_0, x = var_4072_cast_fp16)[name = string("op_4073_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_683_cast_fp16 = mul(x = hidden_states_41_cast_fp16, y = var_4073_cast_fp16)[name = string("x_683_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_20_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_20_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1175306240)))];
            tensor<fp16, [1, 1024, 1024]> var_4076_cast_fp16 = mul(x = x_683_cast_fp16, y = encoder_layers_20_input_layernorm_weight_promoted_to_fp16)[name = string("op_4076_cast_fp16")];
            tensor<int32, [3]> var_4081 = const()[name = string("op_4081"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_201_axes_0 = const()[name = string("input_201_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4082 = transpose(perm = var_4081, x = var_4076_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 1024, 1, 1024]> input_201 = expand_dims(axes = input_201_axes_0, x = var_4082)[name = string("input_201")];
            string var_4089_pad_type_0 = const()[name = string("op_4089_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4089_strides_0 = const()[name = string("op_4089_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4089_pad_0 = const()[name = string("op_4089_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4089_dilations_0 = const()[name = string("op_4089_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4089_groups_0 = const()[name = string("op_4089_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_4089 = conv(dilations = var_4089_dilations_0, groups = var_4089_groups_0, pad = var_4089_pad_0, pad_type = var_4089_pad_type_0, strides = var_4089_strides_0, weight = encoder_layers_20_self_attn_q_proj_weight, x = input_201)[name = string("op_4089")];
            tensor<int32, [4]> var_4090 = const()[name = string("op_4090"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_4091 = reshape(shape = var_4090, x = var_4089)[name = string("op_4091")];
            tensor<int32, [4]> var_4092 = const()[name = string("op_4092"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4099_pad_type_0 = const()[name = string("op_4099_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4099_strides_0 = const()[name = string("op_4099_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4099_pad_0 = const()[name = string("op_4099_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4099_dilations_0 = const()[name = string("op_4099_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4099_groups_0 = const()[name = string("op_4099_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4099 = conv(dilations = var_4099_dilations_0, groups = var_4099_groups_0, pad = var_4099_pad_0, pad_type = var_4099_pad_type_0, strides = var_4099_strides_0, weight = encoder_layers_20_self_attn_k_proj_weight, x = input_201)[name = string("op_4099")];
            tensor<int32, [4]> var_4100 = const()[name = string("op_4100"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_4101 = reshape(shape = var_4100, x = var_4099)[name = string("op_4101")];
            tensor<int32, [4]> var_4102 = const()[name = string("op_4102"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4109_pad_type_0 = const()[name = string("op_4109_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4109_strides_0 = const()[name = string("op_4109_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4109_pad_0 = const()[name = string("op_4109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4109_dilations_0 = const()[name = string("op_4109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4109_groups_0 = const()[name = string("op_4109_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4109 = conv(dilations = var_4109_dilations_0, groups = var_4109_groups_0, pad = var_4109_pad_0, pad_type = var_4109_pad_type_0, strides = var_4109_strides_0, weight = encoder_layers_20_self_attn_v_proj_weight, x = input_201)[name = string("op_4109")];
            tensor<int32, [4]> var_4110 = const()[name = string("op_4110"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_4111 = reshape(shape = var_4110, x = var_4109)[name = string("op_4111")];
            tensor<int32, [4]> var_4112 = const()[name = string("op_4112"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_81_to_fp16 = const()[name = string("op_5_promoted_81_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_121 = transpose(perm = var_4092, x = var_4091)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 1024, 128]> var_4118_cast_fp16 = pow(x = q_121, y = var_5_promoted_81_to_fp16)[name = string("op_4118_cast_fp16")];
            tensor<int32, [1]> var_163_axes_0 = const()[name = string("var_163_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_163_keep_dims_0 = const()[name = string("var_163_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_163_cast_fp16 = reduce_mean(axes = var_163_axes_0, keep_dims = var_163_keep_dims_0, x = var_4118_cast_fp16)[name = string("var_163_cast_fp16")];
            fp16 var_4121_to_fp16 = const()[name = string("op_4121_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_4122_cast_fp16 = add(x = var_163_cast_fp16, y = var_4121_to_fp16)[name = string("op_4122_cast_fp16")];
            fp32 var_4123_epsilon_0 = const()[name = string("op_4123_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_4123_cast_fp16 = rsqrt(epsilon = var_4123_epsilon_0, x = var_4122_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_691_cast_fp16 = mul(x = q_121, y = var_4123_cast_fp16)[name = string("x_691_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_123 = mul(x = x_691_cast_fp16, y = encoder_layers_20_self_attn_q_norm_weight)[name = string("q_123")];
            fp16 var_5_promoted_82_to_fp16 = const()[name = string("op_5_promoted_82_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_121 = transpose(perm = var_4102, x = var_4101)[name = string("transpose_69")];
            tensor<fp16, [1, 8, 1024, 128]> var_4131_cast_fp16 = pow(x = k_121, y = var_5_promoted_82_to_fp16)[name = string("op_4131_cast_fp16")];
            tensor<int32, [1]> var_165_axes_0 = const()[name = string("var_165_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_165_keep_dims_0 = const()[name = string("var_165_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_165_cast_fp16 = reduce_mean(axes = var_165_axes_0, keep_dims = var_165_keep_dims_0, x = var_4131_cast_fp16)[name = string("var_165_cast_fp16")];
            fp16 var_4134_to_fp16 = const()[name = string("op_4134_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_4135_cast_fp16 = add(x = var_165_cast_fp16, y = var_4134_to_fp16)[name = string("op_4135_cast_fp16")];
            fp32 var_4136_epsilon_0 = const()[name = string("op_4136_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_4136_cast_fp16 = rsqrt(epsilon = var_4136_epsilon_0, x = var_4135_cast_fp16)[name = string("op_4136_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_697_cast_fp16 = mul(x = k_121, y = var_4136_cast_fp16)[name = string("x_697_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_123 = mul(x = x_697_cast_fp16, y = encoder_layers_20_self_attn_k_norm_weight)[name = string("k_123")];
            tensor<fp16, [1, 16, 1024, 128]> var_4140 = mul(x = q_123, y = cos)[name = string("op_4140")];
            tensor<int32, [2]> var_4141_split_sizes_0 = const()[name = string("op_4141_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4141_axis_0 = const()[name = string("op_4141_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_4141_0, tensor<fp16, [1, 16, 1024, 64]> var_4141_1 = split(axis = var_4141_axis_0, split_sizes = var_4141_split_sizes_0, x = q_123)[name = string("op_4141")];
            fp16 const_63_promoted = const()[name = string("const_63_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_4143 = mul(x = var_4141_1, y = const_63_promoted)[name = string("op_4143")];
            bool var_4145_interleave_0 = const()[name = string("op_4145_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_4145 = concat(axis = var_17, interleave = var_4145_interleave_0, values = (var_4143, var_4141_0))[name = string("op_4145")];
            tensor<fp16, [1, 16, 1024, 128]> var_4146 = mul(x = var_4145, y = sin)[name = string("op_4146")];
            tensor<fp16, [1, 16, 1024, 128]> query_41 = add(x = var_4140, y = var_4146)[name = string("query_41")];
            tensor<fp16, [1, 8, 1024, 128]> var_4148 = mul(x = k_123, y = cos)[name = string("op_4148")];
            tensor<int32, [2]> var_4149_split_sizes_0 = const()[name = string("op_4149_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4149_axis_0 = const()[name = string("op_4149_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_4149_0, tensor<fp16, [1, 8, 1024, 64]> var_4149_1 = split(axis = var_4149_axis_0, split_sizes = var_4149_split_sizes_0, x = k_123)[name = string("op_4149")];
            fp16 const_64_promoted = const()[name = string("const_64_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_4151 = mul(x = var_4149_1, y = const_64_promoted)[name = string("op_4151")];
            bool var_4153_interleave_0 = const()[name = string("op_4153_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_4153 = concat(axis = var_17, interleave = var_4153_interleave_0, values = (var_4151, var_4149_0))[name = string("op_4153")];
            tensor<fp16, [1, 8, 1024, 128]> var_4154 = mul(x = var_4153, y = sin)[name = string("op_4154")];
            tensor<fp16, [1, 8, 1024, 128]> x_699 = add(x = var_4148, y = var_4154)[name = string("x_699")];
            tensor<int32, [1]> var_4156_axes_0 = const()[name = string("op_4156_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4156 = expand_dims(axes = var_4156_axes_0, x = x_699)[name = string("op_4156")];
            tensor<int32, [5]> x_701_reps_0 = const()[name = string("x_701_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_701 = tile(reps = x_701_reps_0, x = var_4156)[name = string("x_701")];
            tensor<int32, [4]> var_4159 = const()[name = string("op_4159"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_41 = reshape(shape = var_4159, x = x_701)[name = string("key_41")];
            tensor<int32, [1]> var_4161_axes_0 = const()[name = string("op_4161_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_703 = transpose(perm = var_4112, x = var_4111)[name = string("transpose_68")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4161 = expand_dims(axes = var_4161_axes_0, x = x_703)[name = string("op_4161")];
            tensor<int32, [5]> x_705_reps_0 = const()[name = string("x_705_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_705 = tile(reps = x_705_reps_0, x = var_4161)[name = string("x_705")];
            tensor<int32, [4]> var_4164 = const()[name = string("op_4164"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_41 = reshape(shape = var_4164, x = x_705)[name = string("value_41")];
            bool var_4169_transpose_x_1 = const()[name = string("op_4169_transpose_x_1"), val = bool(false)];
            bool var_4169_transpose_y_1 = const()[name = string("op_4169_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_4169_cast_fp16 = matmul(transpose_x = var_4169_transpose_x_1, transpose_y = var_4169_transpose_y_1, x = query_41, y = key_41)[name = string("op_4169_cast_fp16")];
            fp16 var_4170_to_fp16 = const()[name = string("op_4170_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_121_cast_fp16 = mul(x = var_4169_cast_fp16, y = var_4170_to_fp16)[name = string("attn_weights_121_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_123_cast_fp16 = add(x = attn_weights_121_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_4174_cast_fp16 = softmax(axis = var_17, x = attn_weights_123_cast_fp16)[name = string("op_4174_cast_fp16")];
            bool var_4178_transpose_x_0 = const()[name = string("op_4178_transpose_x_0"), val = bool(false)];
            bool var_4178_transpose_y_0 = const()[name = string("op_4178_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_4178_cast_fp16 = matmul(transpose_x = var_4178_transpose_x_0, transpose_y = var_4178_transpose_y_0, x = var_4174_cast_fp16, y = value_41)[name = string("op_4178_cast_fp16")];
            tensor<int32, [4]> var_4180 = const()[name = string("op_4180"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4183 = const()[name = string("op_4183"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_4181 = transpose(perm = var_4180, x = var_4178_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 1024, 2048]> attn_out_123 = reshape(shape = var_4183, x = var_4181)[name = string("attn_out_123")];
            tensor<int32, [3]> var_4185 = const()[name = string("op_4185"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_20 = const()[name = string("squeeze_20"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1175308352)))];
            string var_4194_pad_type_0 = const()[name = string("op_4194_pad_type_0"), val = string("valid")];
            int32 var_4194_groups_0 = const()[name = string("op_4194_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_4194_strides_0 = const()[name = string("op_4194_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_4194_pad_0 = const()[name = string("op_4194_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_4194_dilations_0 = const()[name = string("op_4194_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_4186 = transpose(perm = var_4185, x = attn_out_123)[name = string("transpose_66")];
            tensor<fp16, [1, 1024, 1024]> var_4194 = conv(dilations = var_4194_dilations_0, groups = var_4194_groups_0, pad = var_4194_pad_0, pad_type = var_4194_pad_type_0, strides = var_4194_strides_0, weight = squeeze_20, x = var_4186)[name = string("op_4194")];
            tensor<int32, [3]> var_4195 = const()[name = string("op_4195"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_125 = transpose(perm = var_4195, x = var_4194)[name = string("transpose_65")];
            tensor<fp16, [1, 1024, 1024]> x_707_cast_fp16 = add(x = hidden_states_41_cast_fp16, y = attn_out_125)[name = string("x_707_cast_fp16")];
            fp16 var_5_promoted_83_to_fp16 = const()[name = string("op_5_promoted_83_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_4201_cast_fp16 = pow(x = x_707_cast_fp16, y = var_5_promoted_83_to_fp16)[name = string("op_4201_cast_fp16")];
            tensor<int32, [1]> var_167_axes_0 = const()[name = string("var_167_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_167_keep_dims_0 = const()[name = string("var_167_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_167_cast_fp16 = reduce_mean(axes = var_167_axes_0, keep_dims = var_167_keep_dims_0, x = var_4201_cast_fp16)[name = string("var_167_cast_fp16")];
            fp16 var_4204_to_fp16 = const()[name = string("op_4204_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_4205_cast_fp16 = add(x = var_167_cast_fp16, y = var_4204_to_fp16)[name = string("op_4205_cast_fp16")];
            fp32 var_4206_epsilon_0 = const()[name = string("op_4206_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_4206_cast_fp16 = rsqrt(epsilon = var_4206_epsilon_0, x = var_4205_cast_fp16)[name = string("op_4206_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_711_cast_fp16 = mul(x = x_707_cast_fp16, y = var_4206_cast_fp16)[name = string("x_711_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_20_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_20_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1179502720)))];
            tensor<fp16, [1, 1024, 1024]> var_4209_cast_fp16 = mul(x = x_711_cast_fp16, y = encoder_layers_20_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4209_cast_fp16")];
            tensor<int32, [3]> var_4214 = const()[name = string("op_4214"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_205_axes_0 = const()[name = string("input_205_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4215 = transpose(perm = var_4214, x = var_4209_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 1024, 1, 1024]> input_205 = expand_dims(axes = input_205_axes_0, x = var_4215)[name = string("input_205")];
            string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_207 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_20_mlp_gate_proj_weight, x = input_205)[name = string("input_207")];
            string up_41_pad_type_0 = const()[name = string("up_41_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_41_strides_0 = const()[name = string("up_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_41_pad_0 = const()[name = string("up_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_41_dilations_0 = const()[name = string("up_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_41_groups_0 = const()[name = string("up_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_41 = conv(dilations = up_41_dilations_0, groups = up_41_groups_0, pad = up_41_pad_0, pad_type = up_41_pad_type_0, strides = up_41_strides_0, weight = encoder_layers_20_mlp_up_proj_weight, x = input_205)[name = string("up_41")];
            tensor<fp16, [1, 3072, 1, 1024]> var_4229 = silu(x = input_207)[name = string("op_4229")];
            tensor<fp16, [1, 3072, 1, 1024]> input_209 = mul(x = var_4229, y = up_41)[name = string("input_209")];
            string var_4236_pad_type_0 = const()[name = string("op_4236_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4236_strides_0 = const()[name = string("op_4236_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4236_pad_0 = const()[name = string("op_4236_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4236_dilations_0 = const()[name = string("op_4236_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4236_groups_0 = const()[name = string("op_4236_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4236 = conv(dilations = var_4236_dilations_0, groups = var_4236_groups_0, pad = var_4236_pad_0, pad_type = var_4236_pad_type_0, strides = var_4236_strides_0, weight = encoder_layers_20_mlp_down_proj_weight, x = input_209)[name = string("op_4236")];
            tensor<int32, [1]> var_4237_axes_0 = const()[name = string("op_4237_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4237 = squeeze(axes = var_4237_axes_0, x = var_4236)[name = string("op_4237")];
            tensor<int32, [3]> var_4238 = const()[name = string("op_4238"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_41 = transpose(perm = var_4238, x = var_4237)[name = string("transpose_63")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_43_cast_fp16 = add(x = x_707_cast_fp16, y = mlp_out_41)[name = string("hidden_states_43_cast_fp16")];
            fp16 var_5_promoted_84_to_fp16 = const()[name = string("op_5_promoted_84_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_4265_cast_fp16 = pow(x = hidden_states_43_cast_fp16, y = var_5_promoted_84_to_fp16)[name = string("op_4265_cast_fp16")];
            tensor<int32, [1]> var_169_axes_0 = const()[name = string("var_169_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_169_keep_dims_0 = const()[name = string("var_169_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_169_cast_fp16 = reduce_mean(axes = var_169_axes_0, keep_dims = var_169_keep_dims_0, x = var_4265_cast_fp16)[name = string("var_169_cast_fp16")];
            fp16 var_4268_to_fp16 = const()[name = string("op_4268_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_4269_cast_fp16 = add(x = var_169_cast_fp16, y = var_4268_to_fp16)[name = string("op_4269_cast_fp16")];
            fp32 var_4270_epsilon_0 = const()[name = string("op_4270_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_4270_cast_fp16 = rsqrt(epsilon = var_4270_epsilon_0, x = var_4269_cast_fp16)[name = string("op_4270_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_717_cast_fp16 = mul(x = hidden_states_43_cast_fp16, y = var_4270_cast_fp16)[name = string("x_717_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_21_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_21_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1179504832)))];
            tensor<fp16, [1, 1024, 1024]> var_4273_cast_fp16 = mul(x = x_717_cast_fp16, y = encoder_layers_21_input_layernorm_weight_promoted_to_fp16)[name = string("op_4273_cast_fp16")];
            tensor<int32, [3]> var_4278 = const()[name = string("op_4278"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_211_axes_0 = const()[name = string("input_211_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4279 = transpose(perm = var_4278, x = var_4273_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 1024, 1, 1024]> input_211 = expand_dims(axes = input_211_axes_0, x = var_4279)[name = string("input_211")];
            string var_4286_pad_type_0 = const()[name = string("op_4286_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4286_strides_0 = const()[name = string("op_4286_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4286_pad_0 = const()[name = string("op_4286_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4286_dilations_0 = const()[name = string("op_4286_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4286_groups_0 = const()[name = string("op_4286_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_4286 = conv(dilations = var_4286_dilations_0, groups = var_4286_groups_0, pad = var_4286_pad_0, pad_type = var_4286_pad_type_0, strides = var_4286_strides_0, weight = encoder_layers_21_self_attn_q_proj_weight, x = input_211)[name = string("op_4286")];
            tensor<int32, [4]> var_4287 = const()[name = string("op_4287"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_4288 = reshape(shape = var_4287, x = var_4286)[name = string("op_4288")];
            tensor<int32, [4]> var_4289 = const()[name = string("op_4289"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4296_pad_type_0 = const()[name = string("op_4296_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4296_strides_0 = const()[name = string("op_4296_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4296_pad_0 = const()[name = string("op_4296_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4296_dilations_0 = const()[name = string("op_4296_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4296_groups_0 = const()[name = string("op_4296_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4296 = conv(dilations = var_4296_dilations_0, groups = var_4296_groups_0, pad = var_4296_pad_0, pad_type = var_4296_pad_type_0, strides = var_4296_strides_0, weight = encoder_layers_21_self_attn_k_proj_weight, x = input_211)[name = string("op_4296")];
            tensor<int32, [4]> var_4297 = const()[name = string("op_4297"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_4298 = reshape(shape = var_4297, x = var_4296)[name = string("op_4298")];
            tensor<int32, [4]> var_4299 = const()[name = string("op_4299"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4306_pad_type_0 = const()[name = string("op_4306_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4306_strides_0 = const()[name = string("op_4306_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4306_pad_0 = const()[name = string("op_4306_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4306_dilations_0 = const()[name = string("op_4306_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4306_groups_0 = const()[name = string("op_4306_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4306 = conv(dilations = var_4306_dilations_0, groups = var_4306_groups_0, pad = var_4306_pad_0, pad_type = var_4306_pad_type_0, strides = var_4306_strides_0, weight = encoder_layers_21_self_attn_v_proj_weight, x = input_211)[name = string("op_4306")];
            tensor<int32, [4]> var_4307 = const()[name = string("op_4307"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_4308 = reshape(shape = var_4307, x = var_4306)[name = string("op_4308")];
            tensor<int32, [4]> var_4309 = const()[name = string("op_4309"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_85_to_fp16 = const()[name = string("op_5_promoted_85_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_127 = transpose(perm = var_4289, x = var_4288)[name = string("transpose_61")];
            tensor<fp16, [1, 16, 1024, 128]> var_4315_cast_fp16 = pow(x = q_127, y = var_5_promoted_85_to_fp16)[name = string("op_4315_cast_fp16")];
            tensor<int32, [1]> var_171_axes_0 = const()[name = string("var_171_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_171_keep_dims_0 = const()[name = string("var_171_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_171_cast_fp16 = reduce_mean(axes = var_171_axes_0, keep_dims = var_171_keep_dims_0, x = var_4315_cast_fp16)[name = string("var_171_cast_fp16")];
            fp16 var_4318_to_fp16 = const()[name = string("op_4318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_4319_cast_fp16 = add(x = var_171_cast_fp16, y = var_4318_to_fp16)[name = string("op_4319_cast_fp16")];
            fp32 var_4320_epsilon_0 = const()[name = string("op_4320_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_4320_cast_fp16 = rsqrt(epsilon = var_4320_epsilon_0, x = var_4319_cast_fp16)[name = string("op_4320_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_725_cast_fp16 = mul(x = q_127, y = var_4320_cast_fp16)[name = string("x_725_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_129 = mul(x = x_725_cast_fp16, y = encoder_layers_21_self_attn_q_norm_weight)[name = string("q_129")];
            fp16 var_5_promoted_86_to_fp16 = const()[name = string("op_5_promoted_86_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_127 = transpose(perm = var_4299, x = var_4298)[name = string("transpose_60")];
            tensor<fp16, [1, 8, 1024, 128]> var_4328_cast_fp16 = pow(x = k_127, y = var_5_promoted_86_to_fp16)[name = string("op_4328_cast_fp16")];
            tensor<int32, [1]> var_173_axes_0 = const()[name = string("var_173_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_173_keep_dims_0 = const()[name = string("var_173_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_173_cast_fp16 = reduce_mean(axes = var_173_axes_0, keep_dims = var_173_keep_dims_0, x = var_4328_cast_fp16)[name = string("var_173_cast_fp16")];
            fp16 var_4331_to_fp16 = const()[name = string("op_4331_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_4332_cast_fp16 = add(x = var_173_cast_fp16, y = var_4331_to_fp16)[name = string("op_4332_cast_fp16")];
            fp32 var_4333_epsilon_0 = const()[name = string("op_4333_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_4333_cast_fp16 = rsqrt(epsilon = var_4333_epsilon_0, x = var_4332_cast_fp16)[name = string("op_4333_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_731_cast_fp16 = mul(x = k_127, y = var_4333_cast_fp16)[name = string("x_731_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_129 = mul(x = x_731_cast_fp16, y = encoder_layers_21_self_attn_k_norm_weight)[name = string("k_129")];
            tensor<fp16, [1, 16, 1024, 128]> var_4337 = mul(x = q_129, y = cos)[name = string("op_4337")];
            tensor<int32, [2]> var_4338_split_sizes_0 = const()[name = string("op_4338_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4338_axis_0 = const()[name = string("op_4338_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_4338_0, tensor<fp16, [1, 16, 1024, 64]> var_4338_1 = split(axis = var_4338_axis_0, split_sizes = var_4338_split_sizes_0, x = q_129)[name = string("op_4338")];
            fp16 const_66_promoted = const()[name = string("const_66_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_4340 = mul(x = var_4338_1, y = const_66_promoted)[name = string("op_4340")];
            bool var_4342_interleave_0 = const()[name = string("op_4342_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_4342 = concat(axis = var_17, interleave = var_4342_interleave_0, values = (var_4340, var_4338_0))[name = string("op_4342")];
            tensor<fp16, [1, 16, 1024, 128]> var_4343 = mul(x = var_4342, y = sin)[name = string("op_4343")];
            tensor<fp16, [1, 16, 1024, 128]> query_43 = add(x = var_4337, y = var_4343)[name = string("query_43")];
            tensor<fp16, [1, 8, 1024, 128]> var_4345 = mul(x = k_129, y = cos)[name = string("op_4345")];
            tensor<int32, [2]> var_4346_split_sizes_0 = const()[name = string("op_4346_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4346_axis_0 = const()[name = string("op_4346_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_4346_0, tensor<fp16, [1, 8, 1024, 64]> var_4346_1 = split(axis = var_4346_axis_0, split_sizes = var_4346_split_sizes_0, x = k_129)[name = string("op_4346")];
            fp16 const_67_promoted = const()[name = string("const_67_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_4348 = mul(x = var_4346_1, y = const_67_promoted)[name = string("op_4348")];
            bool var_4350_interleave_0 = const()[name = string("op_4350_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_4350 = concat(axis = var_17, interleave = var_4350_interleave_0, values = (var_4348, var_4346_0))[name = string("op_4350")];
            tensor<fp16, [1, 8, 1024, 128]> var_4351 = mul(x = var_4350, y = sin)[name = string("op_4351")];
            tensor<fp16, [1, 8, 1024, 128]> x_733 = add(x = var_4345, y = var_4351)[name = string("x_733")];
            tensor<int32, [1]> var_4353_axes_0 = const()[name = string("op_4353_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4353 = expand_dims(axes = var_4353_axes_0, x = x_733)[name = string("op_4353")];
            tensor<int32, [5]> x_735_reps_0 = const()[name = string("x_735_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_735 = tile(reps = x_735_reps_0, x = var_4353)[name = string("x_735")];
            tensor<int32, [4]> var_4356 = const()[name = string("op_4356"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_43 = reshape(shape = var_4356, x = x_735)[name = string("key_43")];
            tensor<int32, [1]> var_4358_axes_0 = const()[name = string("op_4358_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_737 = transpose(perm = var_4309, x = var_4308)[name = string("transpose_59")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4358 = expand_dims(axes = var_4358_axes_0, x = x_737)[name = string("op_4358")];
            tensor<int32, [5]> x_739_reps_0 = const()[name = string("x_739_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_739 = tile(reps = x_739_reps_0, x = var_4358)[name = string("x_739")];
            tensor<int32, [4]> var_4361 = const()[name = string("op_4361"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_43 = reshape(shape = var_4361, x = x_739)[name = string("value_43")];
            bool var_4366_transpose_x_1 = const()[name = string("op_4366_transpose_x_1"), val = bool(false)];
            bool var_4366_transpose_y_1 = const()[name = string("op_4366_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_4366_cast_fp16 = matmul(transpose_x = var_4366_transpose_x_1, transpose_y = var_4366_transpose_y_1, x = query_43, y = key_43)[name = string("op_4366_cast_fp16")];
            fp16 var_4367_to_fp16 = const()[name = string("op_4367_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_127_cast_fp16 = mul(x = var_4366_cast_fp16, y = var_4367_to_fp16)[name = string("attn_weights_127_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_129_cast_fp16 = add(x = attn_weights_127_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_129_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_4371_cast_fp16 = softmax(axis = var_17, x = attn_weights_129_cast_fp16)[name = string("op_4371_cast_fp16")];
            bool var_4375_transpose_x_0 = const()[name = string("op_4375_transpose_x_0"), val = bool(false)];
            bool var_4375_transpose_y_0 = const()[name = string("op_4375_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_4375_cast_fp16 = matmul(transpose_x = var_4375_transpose_x_0, transpose_y = var_4375_transpose_y_0, x = var_4371_cast_fp16, y = value_43)[name = string("op_4375_cast_fp16")];
            tensor<int32, [4]> var_4377 = const()[name = string("op_4377"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4380 = const()[name = string("op_4380"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_4378 = transpose(perm = var_4377, x = var_4375_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 1024, 2048]> attn_out_129 = reshape(shape = var_4380, x = var_4378)[name = string("attn_out_129")];
            tensor<int32, [3]> var_4382 = const()[name = string("op_4382"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_21 = const()[name = string("squeeze_21"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1179506944)))];
            string var_4391_pad_type_0 = const()[name = string("op_4391_pad_type_0"), val = string("valid")];
            int32 var_4391_groups_0 = const()[name = string("op_4391_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_4391_strides_0 = const()[name = string("op_4391_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_4391_pad_0 = const()[name = string("op_4391_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_4391_dilations_0 = const()[name = string("op_4391_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_4383 = transpose(perm = var_4382, x = attn_out_129)[name = string("transpose_57")];
            tensor<fp16, [1, 1024, 1024]> var_4391 = conv(dilations = var_4391_dilations_0, groups = var_4391_groups_0, pad = var_4391_pad_0, pad_type = var_4391_pad_type_0, strides = var_4391_strides_0, weight = squeeze_21, x = var_4383)[name = string("op_4391")];
            tensor<int32, [3]> var_4392 = const()[name = string("op_4392"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_131 = transpose(perm = var_4392, x = var_4391)[name = string("transpose_56")];
            tensor<fp16, [1, 1024, 1024]> x_741_cast_fp16 = add(x = hidden_states_43_cast_fp16, y = attn_out_131)[name = string("x_741_cast_fp16")];
            fp16 var_5_promoted_87_to_fp16 = const()[name = string("op_5_promoted_87_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_4398_cast_fp16 = pow(x = x_741_cast_fp16, y = var_5_promoted_87_to_fp16)[name = string("op_4398_cast_fp16")];
            tensor<int32, [1]> var_175_axes_0 = const()[name = string("var_175_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_175_keep_dims_0 = const()[name = string("var_175_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_175_cast_fp16 = reduce_mean(axes = var_175_axes_0, keep_dims = var_175_keep_dims_0, x = var_4398_cast_fp16)[name = string("var_175_cast_fp16")];
            fp16 var_4401_to_fp16 = const()[name = string("op_4401_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_4402_cast_fp16 = add(x = var_175_cast_fp16, y = var_4401_to_fp16)[name = string("op_4402_cast_fp16")];
            fp32 var_4403_epsilon_0 = const()[name = string("op_4403_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_4403_cast_fp16 = rsqrt(epsilon = var_4403_epsilon_0, x = var_4402_cast_fp16)[name = string("op_4403_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_745_cast_fp16 = mul(x = x_741_cast_fp16, y = var_4403_cast_fp16)[name = string("x_745_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_21_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_21_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1183701312)))];
            tensor<fp16, [1, 1024, 1024]> var_4406_cast_fp16 = mul(x = x_745_cast_fp16, y = encoder_layers_21_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4406_cast_fp16")];
            tensor<int32, [3]> var_4411 = const()[name = string("op_4411"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_215_axes_0 = const()[name = string("input_215_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4412 = transpose(perm = var_4411, x = var_4406_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 1024, 1, 1024]> input_215 = expand_dims(axes = input_215_axes_0, x = var_4412)[name = string("input_215")];
            string input_217_pad_type_0 = const()[name = string("input_217_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_217_strides_0 = const()[name = string("input_217_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_217_pad_0 = const()[name = string("input_217_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_217_dilations_0 = const()[name = string("input_217_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_217_groups_0 = const()[name = string("input_217_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_217 = conv(dilations = input_217_dilations_0, groups = input_217_groups_0, pad = input_217_pad_0, pad_type = input_217_pad_type_0, strides = input_217_strides_0, weight = encoder_layers_21_mlp_gate_proj_weight, x = input_215)[name = string("input_217")];
            string up_43_pad_type_0 = const()[name = string("up_43_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_43_strides_0 = const()[name = string("up_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_43_pad_0 = const()[name = string("up_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_43_dilations_0 = const()[name = string("up_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_43_groups_0 = const()[name = string("up_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_43 = conv(dilations = up_43_dilations_0, groups = up_43_groups_0, pad = up_43_pad_0, pad_type = up_43_pad_type_0, strides = up_43_strides_0, weight = encoder_layers_21_mlp_up_proj_weight, x = input_215)[name = string("up_43")];
            tensor<fp16, [1, 3072, 1, 1024]> var_4426 = silu(x = input_217)[name = string("op_4426")];
            tensor<fp16, [1, 3072, 1, 1024]> input_219 = mul(x = var_4426, y = up_43)[name = string("input_219")];
            string var_4433_pad_type_0 = const()[name = string("op_4433_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4433_strides_0 = const()[name = string("op_4433_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4433_pad_0 = const()[name = string("op_4433_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4433_dilations_0 = const()[name = string("op_4433_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4433_groups_0 = const()[name = string("op_4433_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4433 = conv(dilations = var_4433_dilations_0, groups = var_4433_groups_0, pad = var_4433_pad_0, pad_type = var_4433_pad_type_0, strides = var_4433_strides_0, weight = encoder_layers_21_mlp_down_proj_weight, x = input_219)[name = string("op_4433")];
            tensor<int32, [1]> var_4434_axes_0 = const()[name = string("op_4434_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4434 = squeeze(axes = var_4434_axes_0, x = var_4433)[name = string("op_4434")];
            tensor<int32, [3]> var_4435 = const()[name = string("op_4435"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_43 = transpose(perm = var_4435, x = var_4434)[name = string("transpose_54")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_45_cast_fp16 = add(x = x_741_cast_fp16, y = mlp_out_43)[name = string("hidden_states_45_cast_fp16")];
            fp16 var_5_promoted_88_to_fp16 = const()[name = string("op_5_promoted_88_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_4462_cast_fp16 = pow(x = hidden_states_45_cast_fp16, y = var_5_promoted_88_to_fp16)[name = string("op_4462_cast_fp16")];
            tensor<int32, [1]> var_177_axes_0 = const()[name = string("var_177_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_177_keep_dims_0 = const()[name = string("var_177_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_177_cast_fp16 = reduce_mean(axes = var_177_axes_0, keep_dims = var_177_keep_dims_0, x = var_4462_cast_fp16)[name = string("var_177_cast_fp16")];
            fp16 var_4465_to_fp16 = const()[name = string("op_4465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_4466_cast_fp16 = add(x = var_177_cast_fp16, y = var_4465_to_fp16)[name = string("op_4466_cast_fp16")];
            fp32 var_4467_epsilon_0 = const()[name = string("op_4467_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_4467_cast_fp16 = rsqrt(epsilon = var_4467_epsilon_0, x = var_4466_cast_fp16)[name = string("op_4467_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_751_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = var_4467_cast_fp16)[name = string("x_751_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_22_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_22_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1183703424)))];
            tensor<fp16, [1, 1024, 1024]> var_4470_cast_fp16 = mul(x = x_751_cast_fp16, y = encoder_layers_22_input_layernorm_weight_promoted_to_fp16)[name = string("op_4470_cast_fp16")];
            tensor<int32, [3]> var_4475 = const()[name = string("op_4475"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4476 = transpose(perm = var_4475, x = var_4470_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 1024, 1, 1024]> input_221 = expand_dims(axes = input_221_axes_0, x = var_4476)[name = string("input_221")];
            string var_4483_pad_type_0 = const()[name = string("op_4483_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4483_strides_0 = const()[name = string("op_4483_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4483_pad_0 = const()[name = string("op_4483_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4483_dilations_0 = const()[name = string("op_4483_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4483_groups_0 = const()[name = string("op_4483_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_4483 = conv(dilations = var_4483_dilations_0, groups = var_4483_groups_0, pad = var_4483_pad_0, pad_type = var_4483_pad_type_0, strides = var_4483_strides_0, weight = encoder_layers_22_self_attn_q_proj_weight, x = input_221)[name = string("op_4483")];
            tensor<int32, [4]> var_4484 = const()[name = string("op_4484"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_4485 = reshape(shape = var_4484, x = var_4483)[name = string("op_4485")];
            tensor<int32, [4]> var_4486 = const()[name = string("op_4486"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4493_pad_type_0 = const()[name = string("op_4493_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4493_strides_0 = const()[name = string("op_4493_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4493_pad_0 = const()[name = string("op_4493_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4493_dilations_0 = const()[name = string("op_4493_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4493_groups_0 = const()[name = string("op_4493_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4493 = conv(dilations = var_4493_dilations_0, groups = var_4493_groups_0, pad = var_4493_pad_0, pad_type = var_4493_pad_type_0, strides = var_4493_strides_0, weight = encoder_layers_22_self_attn_k_proj_weight, x = input_221)[name = string("op_4493")];
            tensor<int32, [4]> var_4494 = const()[name = string("op_4494"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_4495 = reshape(shape = var_4494, x = var_4493)[name = string("op_4495")];
            tensor<int32, [4]> var_4496 = const()[name = string("op_4496"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4503_pad_type_0 = const()[name = string("op_4503_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4503_strides_0 = const()[name = string("op_4503_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4503_pad_0 = const()[name = string("op_4503_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4503_dilations_0 = const()[name = string("op_4503_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4503_groups_0 = const()[name = string("op_4503_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4503 = conv(dilations = var_4503_dilations_0, groups = var_4503_groups_0, pad = var_4503_pad_0, pad_type = var_4503_pad_type_0, strides = var_4503_strides_0, weight = encoder_layers_22_self_attn_v_proj_weight, x = input_221)[name = string("op_4503")];
            tensor<int32, [4]> var_4504 = const()[name = string("op_4504"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_4505 = reshape(shape = var_4504, x = var_4503)[name = string("op_4505")];
            tensor<int32, [4]> var_4506 = const()[name = string("op_4506"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_89_to_fp16 = const()[name = string("op_5_promoted_89_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_133 = transpose(perm = var_4486, x = var_4485)[name = string("transpose_52")];
            tensor<fp16, [1, 16, 1024, 128]> var_4512_cast_fp16 = pow(x = q_133, y = var_5_promoted_89_to_fp16)[name = string("op_4512_cast_fp16")];
            tensor<int32, [1]> var_179_axes_0 = const()[name = string("var_179_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_179_keep_dims_0 = const()[name = string("var_179_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_179_cast_fp16 = reduce_mean(axes = var_179_axes_0, keep_dims = var_179_keep_dims_0, x = var_4512_cast_fp16)[name = string("var_179_cast_fp16")];
            fp16 var_4515_to_fp16 = const()[name = string("op_4515_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_4516_cast_fp16 = add(x = var_179_cast_fp16, y = var_4515_to_fp16)[name = string("op_4516_cast_fp16")];
            fp32 var_4517_epsilon_0 = const()[name = string("op_4517_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_4517_cast_fp16 = rsqrt(epsilon = var_4517_epsilon_0, x = var_4516_cast_fp16)[name = string("op_4517_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_759_cast_fp16 = mul(x = q_133, y = var_4517_cast_fp16)[name = string("x_759_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_135 = mul(x = x_759_cast_fp16, y = encoder_layers_22_self_attn_q_norm_weight)[name = string("q_135")];
            fp16 var_5_promoted_90_to_fp16 = const()[name = string("op_5_promoted_90_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_133 = transpose(perm = var_4496, x = var_4495)[name = string("transpose_51")];
            tensor<fp16, [1, 8, 1024, 128]> var_4525_cast_fp16 = pow(x = k_133, y = var_5_promoted_90_to_fp16)[name = string("op_4525_cast_fp16")];
            tensor<int32, [1]> var_181_axes_0 = const()[name = string("var_181_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_181_keep_dims_0 = const()[name = string("var_181_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_181_cast_fp16 = reduce_mean(axes = var_181_axes_0, keep_dims = var_181_keep_dims_0, x = var_4525_cast_fp16)[name = string("var_181_cast_fp16")];
            fp16 var_4528_to_fp16 = const()[name = string("op_4528_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_4529_cast_fp16 = add(x = var_181_cast_fp16, y = var_4528_to_fp16)[name = string("op_4529_cast_fp16")];
            fp32 var_4530_epsilon_0 = const()[name = string("op_4530_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_4530_cast_fp16 = rsqrt(epsilon = var_4530_epsilon_0, x = var_4529_cast_fp16)[name = string("op_4530_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_765_cast_fp16 = mul(x = k_133, y = var_4530_cast_fp16)[name = string("x_765_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_135 = mul(x = x_765_cast_fp16, y = encoder_layers_22_self_attn_k_norm_weight)[name = string("k_135")];
            tensor<fp16, [1, 16, 1024, 128]> var_4534 = mul(x = q_135, y = cos)[name = string("op_4534")];
            tensor<int32, [2]> var_4535_split_sizes_0 = const()[name = string("op_4535_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4535_axis_0 = const()[name = string("op_4535_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_4535_0, tensor<fp16, [1, 16, 1024, 64]> var_4535_1 = split(axis = var_4535_axis_0, split_sizes = var_4535_split_sizes_0, x = q_135)[name = string("op_4535")];
            fp16 const_69_promoted = const()[name = string("const_69_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_4537 = mul(x = var_4535_1, y = const_69_promoted)[name = string("op_4537")];
            bool var_4539_interleave_0 = const()[name = string("op_4539_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_4539 = concat(axis = var_17, interleave = var_4539_interleave_0, values = (var_4537, var_4535_0))[name = string("op_4539")];
            tensor<fp16, [1, 16, 1024, 128]> var_4540 = mul(x = var_4539, y = sin)[name = string("op_4540")];
            tensor<fp16, [1, 16, 1024, 128]> query_45 = add(x = var_4534, y = var_4540)[name = string("query_45")];
            tensor<fp16, [1, 8, 1024, 128]> var_4542 = mul(x = k_135, y = cos)[name = string("op_4542")];
            tensor<int32, [2]> var_4543_split_sizes_0 = const()[name = string("op_4543_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4543_axis_0 = const()[name = string("op_4543_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_4543_0, tensor<fp16, [1, 8, 1024, 64]> var_4543_1 = split(axis = var_4543_axis_0, split_sizes = var_4543_split_sizes_0, x = k_135)[name = string("op_4543")];
            fp16 const_70_promoted = const()[name = string("const_70_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_4545 = mul(x = var_4543_1, y = const_70_promoted)[name = string("op_4545")];
            bool var_4547_interleave_0 = const()[name = string("op_4547_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_4547 = concat(axis = var_17, interleave = var_4547_interleave_0, values = (var_4545, var_4543_0))[name = string("op_4547")];
            tensor<fp16, [1, 8, 1024, 128]> var_4548 = mul(x = var_4547, y = sin)[name = string("op_4548")];
            tensor<fp16, [1, 8, 1024, 128]> x_767 = add(x = var_4542, y = var_4548)[name = string("x_767")];
            tensor<int32, [1]> var_4550_axes_0 = const()[name = string("op_4550_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4550 = expand_dims(axes = var_4550_axes_0, x = x_767)[name = string("op_4550")];
            tensor<int32, [5]> x_769_reps_0 = const()[name = string("x_769_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_769 = tile(reps = x_769_reps_0, x = var_4550)[name = string("x_769")];
            tensor<int32, [4]> var_4553 = const()[name = string("op_4553"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_45 = reshape(shape = var_4553, x = x_769)[name = string("key_45")];
            tensor<int32, [1]> var_4555_axes_0 = const()[name = string("op_4555_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_771 = transpose(perm = var_4506, x = var_4505)[name = string("transpose_50")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4555 = expand_dims(axes = var_4555_axes_0, x = x_771)[name = string("op_4555")];
            tensor<int32, [5]> x_773_reps_0 = const()[name = string("x_773_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_773 = tile(reps = x_773_reps_0, x = var_4555)[name = string("x_773")];
            tensor<int32, [4]> var_4558 = const()[name = string("op_4558"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_45 = reshape(shape = var_4558, x = x_773)[name = string("value_45")];
            bool var_4563_transpose_x_1 = const()[name = string("op_4563_transpose_x_1"), val = bool(false)];
            bool var_4563_transpose_y_1 = const()[name = string("op_4563_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_4563_cast_fp16 = matmul(transpose_x = var_4563_transpose_x_1, transpose_y = var_4563_transpose_y_1, x = query_45, y = key_45)[name = string("op_4563_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_133_cast_fp16 = mul(x = var_4563_cast_fp16, y = var_4564_to_fp16)[name = string("attn_weights_133_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_135_cast_fp16 = add(x = attn_weights_133_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_4568_cast_fp16 = softmax(axis = var_17, x = attn_weights_135_cast_fp16)[name = string("op_4568_cast_fp16")];
            bool var_4572_transpose_x_0 = const()[name = string("op_4572_transpose_x_0"), val = bool(false)];
            bool var_4572_transpose_y_0 = const()[name = string("op_4572_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_4572_cast_fp16 = matmul(transpose_x = var_4572_transpose_x_0, transpose_y = var_4572_transpose_y_0, x = var_4568_cast_fp16, y = value_45)[name = string("op_4572_cast_fp16")];
            tensor<int32, [4]> var_4574 = const()[name = string("op_4574"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4577 = const()[name = string("op_4577"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_4575 = transpose(perm = var_4574, x = var_4572_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 1024, 2048]> attn_out_135 = reshape(shape = var_4577, x = var_4575)[name = string("attn_out_135")];
            tensor<int32, [3]> var_4579 = const()[name = string("op_4579"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_22 = const()[name = string("squeeze_22"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1183705536)))];
            string var_4588_pad_type_0 = const()[name = string("op_4588_pad_type_0"), val = string("valid")];
            int32 var_4588_groups_0 = const()[name = string("op_4588_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_4588_strides_0 = const()[name = string("op_4588_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_4588_pad_0 = const()[name = string("op_4588_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_4588_dilations_0 = const()[name = string("op_4588_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_4580 = transpose(perm = var_4579, x = attn_out_135)[name = string("transpose_48")];
            tensor<fp16, [1, 1024, 1024]> var_4588 = conv(dilations = var_4588_dilations_0, groups = var_4588_groups_0, pad = var_4588_pad_0, pad_type = var_4588_pad_type_0, strides = var_4588_strides_0, weight = squeeze_22, x = var_4580)[name = string("op_4588")];
            tensor<int32, [3]> var_4589 = const()[name = string("op_4589"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_137 = transpose(perm = var_4589, x = var_4588)[name = string("transpose_47")];
            tensor<fp16, [1, 1024, 1024]> x_775_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = attn_out_137)[name = string("x_775_cast_fp16")];
            fp16 var_5_promoted_91_to_fp16 = const()[name = string("op_5_promoted_91_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_4595_cast_fp16 = pow(x = x_775_cast_fp16, y = var_5_promoted_91_to_fp16)[name = string("op_4595_cast_fp16")];
            tensor<int32, [1]> var_183_axes_0 = const()[name = string("var_183_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_183_keep_dims_0 = const()[name = string("var_183_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_183_cast_fp16_0 = reduce_mean(axes = var_183_axes_0, keep_dims = var_183_keep_dims_0, x = var_4595_cast_fp16)[name = string("var_183_cast_fp16")];
            fp16 var_4598_to_fp16 = const()[name = string("op_4598_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_4599_cast_fp16 = add(x = var_183_cast_fp16_0, y = var_4598_to_fp16)[name = string("op_4599_cast_fp16")];
            fp32 var_4600_epsilon_0 = const()[name = string("op_4600_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_4600_cast_fp16 = rsqrt(epsilon = var_4600_epsilon_0, x = var_4599_cast_fp16)[name = string("op_4600_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_779_cast_fp16 = mul(x = x_775_cast_fp16, y = var_4600_cast_fp16)[name = string("x_779_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_22_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_22_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1187899904)))];
            tensor<fp16, [1, 1024, 1024]> var_4603_cast_fp16 = mul(x = x_779_cast_fp16, y = encoder_layers_22_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4603_cast_fp16")];
            tensor<int32, [3]> var_4608 = const()[name = string("op_4608"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_225_axes_0 = const()[name = string("input_225_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4609 = transpose(perm = var_4608, x = var_4603_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 1024, 1, 1024]> input_225 = expand_dims(axes = input_225_axes_0, x = var_4609)[name = string("input_225")];
            string input_227_pad_type_0 = const()[name = string("input_227_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_227_strides_0 = const()[name = string("input_227_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_227_pad_0 = const()[name = string("input_227_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_227_dilations_0 = const()[name = string("input_227_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_227_groups_0 = const()[name = string("input_227_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_227 = conv(dilations = input_227_dilations_0, groups = input_227_groups_0, pad = input_227_pad_0, pad_type = input_227_pad_type_0, strides = input_227_strides_0, weight = encoder_layers_22_mlp_gate_proj_weight, x = input_225)[name = string("input_227")];
            string up_45_pad_type_0 = const()[name = string("up_45_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_45_strides_0 = const()[name = string("up_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_45_pad_0 = const()[name = string("up_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_45_dilations_0 = const()[name = string("up_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_45_groups_0 = const()[name = string("up_45_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_45 = conv(dilations = up_45_dilations_0, groups = up_45_groups_0, pad = up_45_pad_0, pad_type = up_45_pad_type_0, strides = up_45_strides_0, weight = encoder_layers_22_mlp_up_proj_weight, x = input_225)[name = string("up_45")];
            tensor<fp16, [1, 3072, 1, 1024]> var_4623 = silu(x = input_227)[name = string("op_4623")];
            tensor<fp16, [1, 3072, 1, 1024]> input_229 = mul(x = var_4623, y = up_45)[name = string("input_229")];
            string var_4630_pad_type_0 = const()[name = string("op_4630_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4630_strides_0 = const()[name = string("op_4630_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4630_pad_0 = const()[name = string("op_4630_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4630_dilations_0 = const()[name = string("op_4630_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4630_groups_0 = const()[name = string("op_4630_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4630 = conv(dilations = var_4630_dilations_0, groups = var_4630_groups_0, pad = var_4630_pad_0, pad_type = var_4630_pad_type_0, strides = var_4630_strides_0, weight = encoder_layers_22_mlp_down_proj_weight, x = input_229)[name = string("op_4630")];
            tensor<int32, [1]> var_4631_axes_0 = const()[name = string("op_4631_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4631 = squeeze(axes = var_4631_axes_0, x = var_4630)[name = string("op_4631")];
            tensor<int32, [3]> var_4632 = const()[name = string("op_4632"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_45 = transpose(perm = var_4632, x = var_4631)[name = string("transpose_45")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_47_cast_fp16 = add(x = x_775_cast_fp16, y = mlp_out_45)[name = string("hidden_states_47_cast_fp16")];
            fp16 var_5_promoted_92_to_fp16 = const()[name = string("op_5_promoted_92_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_4659_cast_fp16 = pow(x = hidden_states_47_cast_fp16, y = var_5_promoted_92_to_fp16)[name = string("op_4659_cast_fp16")];
            tensor<int32, [1]> var_185_axes_0 = const()[name = string("var_185_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_185_keep_dims_0 = const()[name = string("var_185_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_185_cast_fp16 = reduce_mean(axes = var_185_axes_0, keep_dims = var_185_keep_dims_0, x = var_4659_cast_fp16)[name = string("var_185_cast_fp16")];
            fp16 var_4662_to_fp16 = const()[name = string("op_4662_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_4663_cast_fp16 = add(x = var_185_cast_fp16, y = var_4662_to_fp16)[name = string("op_4663_cast_fp16")];
            fp32 var_4664_epsilon_0 = const()[name = string("op_4664_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_4664_cast_fp16 = rsqrt(epsilon = var_4664_epsilon_0, x = var_4663_cast_fp16)[name = string("op_4664_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_785_cast_fp16 = mul(x = hidden_states_47_cast_fp16, y = var_4664_cast_fp16)[name = string("x_785_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_23_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_23_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1187902016)))];
            tensor<fp16, [1, 1024, 1024]> var_4667_cast_fp16 = mul(x = x_785_cast_fp16, y = encoder_layers_23_input_layernorm_weight_promoted_to_fp16)[name = string("op_4667_cast_fp16")];
            tensor<int32, [3]> var_4672 = const()[name = string("op_4672"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_231_axes_0 = const()[name = string("input_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4673 = transpose(perm = var_4672, x = var_4667_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 1024, 1, 1024]> input_231 = expand_dims(axes = input_231_axes_0, x = var_4673)[name = string("input_231")];
            string var_4680_pad_type_0 = const()[name = string("op_4680_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4680_strides_0 = const()[name = string("op_4680_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4680_pad_0 = const()[name = string("op_4680_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4680_dilations_0 = const()[name = string("op_4680_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4680_groups_0 = const()[name = string("op_4680_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_4680 = conv(dilations = var_4680_dilations_0, groups = var_4680_groups_0, pad = var_4680_pad_0, pad_type = var_4680_pad_type_0, strides = var_4680_strides_0, weight = encoder_layers_23_self_attn_q_proj_weight, x = input_231)[name = string("op_4680")];
            tensor<int32, [4]> var_4681 = const()[name = string("op_4681"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_4682 = reshape(shape = var_4681, x = var_4680)[name = string("op_4682")];
            tensor<int32, [4]> var_4683 = const()[name = string("op_4683"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4690_pad_type_0 = const()[name = string("op_4690_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4690_strides_0 = const()[name = string("op_4690_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4690_pad_0 = const()[name = string("op_4690_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4690_dilations_0 = const()[name = string("op_4690_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4690_groups_0 = const()[name = string("op_4690_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4690 = conv(dilations = var_4690_dilations_0, groups = var_4690_groups_0, pad = var_4690_pad_0, pad_type = var_4690_pad_type_0, strides = var_4690_strides_0, weight = encoder_layers_23_self_attn_k_proj_weight, x = input_231)[name = string("op_4690")];
            tensor<int32, [4]> var_4691 = const()[name = string("op_4691"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_4692 = reshape(shape = var_4691, x = var_4690)[name = string("op_4692")];
            tensor<int32, [4]> var_4693 = const()[name = string("op_4693"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4700_pad_type_0 = const()[name = string("op_4700_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4700_strides_0 = const()[name = string("op_4700_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4700_pad_0 = const()[name = string("op_4700_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4700_dilations_0 = const()[name = string("op_4700_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4700_groups_0 = const()[name = string("op_4700_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4700 = conv(dilations = var_4700_dilations_0, groups = var_4700_groups_0, pad = var_4700_pad_0, pad_type = var_4700_pad_type_0, strides = var_4700_strides_0, weight = encoder_layers_23_self_attn_v_proj_weight, x = input_231)[name = string("op_4700")];
            tensor<int32, [4]> var_4701 = const()[name = string("op_4701"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_4702 = reshape(shape = var_4701, x = var_4700)[name = string("op_4702")];
            tensor<int32, [4]> var_4703 = const()[name = string("op_4703"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_93_to_fp16 = const()[name = string("op_5_promoted_93_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_139 = transpose(perm = var_4683, x = var_4682)[name = string("transpose_43")];
            tensor<fp16, [1, 16, 1024, 128]> var_4709_cast_fp16 = pow(x = q_139, y = var_5_promoted_93_to_fp16)[name = string("op_4709_cast_fp16")];
            tensor<int32, [1]> var_187_axes_0 = const()[name = string("var_187_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_187_keep_dims_0 = const()[name = string("var_187_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_187_cast_fp16 = reduce_mean(axes = var_187_axes_0, keep_dims = var_187_keep_dims_0, x = var_4709_cast_fp16)[name = string("var_187_cast_fp16")];
            fp16 var_4712_to_fp16 = const()[name = string("op_4712_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_4713_cast_fp16 = add(x = var_187_cast_fp16, y = var_4712_to_fp16)[name = string("op_4713_cast_fp16")];
            fp32 var_4714_epsilon_0 = const()[name = string("op_4714_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_4714_cast_fp16 = rsqrt(epsilon = var_4714_epsilon_0, x = var_4713_cast_fp16)[name = string("op_4714_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_793_cast_fp16 = mul(x = q_139, y = var_4714_cast_fp16)[name = string("x_793_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_141 = mul(x = x_793_cast_fp16, y = encoder_layers_23_self_attn_q_norm_weight)[name = string("q_141")];
            fp16 var_5_promoted_94_to_fp16 = const()[name = string("op_5_promoted_94_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_139 = transpose(perm = var_4693, x = var_4692)[name = string("transpose_42")];
            tensor<fp16, [1, 8, 1024, 128]> var_4722_cast_fp16 = pow(x = k_139, y = var_5_promoted_94_to_fp16)[name = string("op_4722_cast_fp16")];
            tensor<int32, [1]> var_189_axes_0 = const()[name = string("var_189_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_189_keep_dims_0 = const()[name = string("var_189_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_189_cast_fp16 = reduce_mean(axes = var_189_axes_0, keep_dims = var_189_keep_dims_0, x = var_4722_cast_fp16)[name = string("var_189_cast_fp16")];
            fp16 var_4725_to_fp16 = const()[name = string("op_4725_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_4726_cast_fp16 = add(x = var_189_cast_fp16, y = var_4725_to_fp16)[name = string("op_4726_cast_fp16")];
            fp32 var_4727_epsilon_0 = const()[name = string("op_4727_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_4727_cast_fp16 = rsqrt(epsilon = var_4727_epsilon_0, x = var_4726_cast_fp16)[name = string("op_4727_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_799_cast_fp16 = mul(x = k_139, y = var_4727_cast_fp16)[name = string("x_799_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_141 = mul(x = x_799_cast_fp16, y = encoder_layers_23_self_attn_k_norm_weight)[name = string("k_141")];
            tensor<fp16, [1, 16, 1024, 128]> var_4731 = mul(x = q_141, y = cos)[name = string("op_4731")];
            tensor<int32, [2]> var_4732_split_sizes_0 = const()[name = string("op_4732_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4732_axis_0 = const()[name = string("op_4732_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_4732_0, tensor<fp16, [1, 16, 1024, 64]> var_4732_1 = split(axis = var_4732_axis_0, split_sizes = var_4732_split_sizes_0, x = q_141)[name = string("op_4732")];
            fp16 const_72_promoted = const()[name = string("const_72_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_4734 = mul(x = var_4732_1, y = const_72_promoted)[name = string("op_4734")];
            bool var_4736_interleave_0 = const()[name = string("op_4736_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_4736 = concat(axis = var_17, interleave = var_4736_interleave_0, values = (var_4734, var_4732_0))[name = string("op_4736")];
            tensor<fp16, [1, 16, 1024, 128]> var_4737 = mul(x = var_4736, y = sin)[name = string("op_4737")];
            tensor<fp16, [1, 16, 1024, 128]> query_47 = add(x = var_4731, y = var_4737)[name = string("query_47")];
            tensor<fp16, [1, 8, 1024, 128]> var_4739 = mul(x = k_141, y = cos)[name = string("op_4739")];
            tensor<int32, [2]> var_4740_split_sizes_0 = const()[name = string("op_4740_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4740_axis_0 = const()[name = string("op_4740_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_4740_0, tensor<fp16, [1, 8, 1024, 64]> var_4740_1 = split(axis = var_4740_axis_0, split_sizes = var_4740_split_sizes_0, x = k_141)[name = string("op_4740")];
            fp16 const_73_promoted = const()[name = string("const_73_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_4742 = mul(x = var_4740_1, y = const_73_promoted)[name = string("op_4742")];
            bool var_4744_interleave_0 = const()[name = string("op_4744_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_4744 = concat(axis = var_17, interleave = var_4744_interleave_0, values = (var_4742, var_4740_0))[name = string("op_4744")];
            tensor<fp16, [1, 8, 1024, 128]> var_4745 = mul(x = var_4744, y = sin)[name = string("op_4745")];
            tensor<fp16, [1, 8, 1024, 128]> x_801 = add(x = var_4739, y = var_4745)[name = string("x_801")];
            tensor<int32, [1]> var_4747_axes_0 = const()[name = string("op_4747_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4747 = expand_dims(axes = var_4747_axes_0, x = x_801)[name = string("op_4747")];
            tensor<int32, [5]> x_803_reps_0 = const()[name = string("x_803_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_803 = tile(reps = x_803_reps_0, x = var_4747)[name = string("x_803")];
            tensor<int32, [4]> var_4750 = const()[name = string("op_4750"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_47 = reshape(shape = var_4750, x = x_803)[name = string("key_47")];
            tensor<int32, [1]> var_4752_axes_0 = const()[name = string("op_4752_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_805 = transpose(perm = var_4703, x = var_4702)[name = string("transpose_41")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4752 = expand_dims(axes = var_4752_axes_0, x = x_805)[name = string("op_4752")];
            tensor<int32, [5]> x_807_reps_0 = const()[name = string("x_807_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_807 = tile(reps = x_807_reps_0, x = var_4752)[name = string("x_807")];
            tensor<int32, [4]> var_4755 = const()[name = string("op_4755"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_47 = reshape(shape = var_4755, x = x_807)[name = string("value_47")];
            bool var_4760_transpose_x_1 = const()[name = string("op_4760_transpose_x_1"), val = bool(false)];
            bool var_4760_transpose_y_1 = const()[name = string("op_4760_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_4760_cast_fp16 = matmul(transpose_x = var_4760_transpose_x_1, transpose_y = var_4760_transpose_y_1, x = query_47, y = key_47)[name = string("op_4760_cast_fp16")];
            fp16 var_4761_to_fp16 = const()[name = string("op_4761_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_139_cast_fp16 = mul(x = var_4760_cast_fp16, y = var_4761_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_4765_cast_fp16 = softmax(axis = var_17, x = attn_weights_141_cast_fp16)[name = string("op_4765_cast_fp16")];
            bool var_4769_transpose_x_0 = const()[name = string("op_4769_transpose_x_0"), val = bool(false)];
            bool var_4769_transpose_y_0 = const()[name = string("op_4769_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_4769_cast_fp16 = matmul(transpose_x = var_4769_transpose_x_0, transpose_y = var_4769_transpose_y_0, x = var_4765_cast_fp16, y = value_47)[name = string("op_4769_cast_fp16")];
            tensor<int32, [4]> var_4771 = const()[name = string("op_4771"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4774 = const()[name = string("op_4774"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_4772 = transpose(perm = var_4771, x = var_4769_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 1024, 2048]> attn_out_141 = reshape(shape = var_4774, x = var_4772)[name = string("attn_out_141")];
            tensor<int32, [3]> var_4776 = const()[name = string("op_4776"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_23 = const()[name = string("squeeze_23"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1187904128)))];
            string var_4785_pad_type_0 = const()[name = string("op_4785_pad_type_0"), val = string("valid")];
            int32 var_4785_groups_0 = const()[name = string("op_4785_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_4785_strides_0 = const()[name = string("op_4785_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_4785_pad_0 = const()[name = string("op_4785_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_4785_dilations_0 = const()[name = string("op_4785_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_4777 = transpose(perm = var_4776, x = attn_out_141)[name = string("transpose_39")];
            tensor<fp16, [1, 1024, 1024]> var_4785 = conv(dilations = var_4785_dilations_0, groups = var_4785_groups_0, pad = var_4785_pad_0, pad_type = var_4785_pad_type_0, strides = var_4785_strides_0, weight = squeeze_23, x = var_4777)[name = string("op_4785")];
            tensor<int32, [3]> var_4786 = const()[name = string("op_4786"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_143 = transpose(perm = var_4786, x = var_4785)[name = string("transpose_38")];
            tensor<fp16, [1, 1024, 1024]> x_809_cast_fp16 = add(x = hidden_states_47_cast_fp16, y = attn_out_143)[name = string("x_809_cast_fp16")];
            fp16 var_5_promoted_95_to_fp16 = const()[name = string("op_5_promoted_95_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_4792_cast_fp16 = pow(x = x_809_cast_fp16, y = var_5_promoted_95_to_fp16)[name = string("op_4792_cast_fp16")];
            tensor<int32, [1]> var_191_axes_0 = const()[name = string("var_191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_191_keep_dims_0 = const()[name = string("var_191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_191_cast_fp16_0 = reduce_mean(axes = var_191_axes_0, keep_dims = var_191_keep_dims_0, x = var_4792_cast_fp16)[name = string("var_191_cast_fp16")];
            fp16 var_4795_to_fp16 = const()[name = string("op_4795_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_4796_cast_fp16 = add(x = var_191_cast_fp16_0, y = var_4795_to_fp16)[name = string("op_4796_cast_fp16")];
            fp32 var_4797_epsilon_0 = const()[name = string("op_4797_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_4797_cast_fp16 = rsqrt(epsilon = var_4797_epsilon_0, x = var_4796_cast_fp16)[name = string("op_4797_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_813_cast_fp16 = mul(x = x_809_cast_fp16, y = var_4797_cast_fp16)[name = string("x_813_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_23_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_23_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1192098496)))];
            tensor<fp16, [1, 1024, 1024]> var_4800_cast_fp16 = mul(x = x_813_cast_fp16, y = encoder_layers_23_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4800_cast_fp16")];
            tensor<int32, [3]> var_4805 = const()[name = string("op_4805"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4806 = transpose(perm = var_4805, x = var_4800_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 1024, 1, 1024]> input_235 = expand_dims(axes = input_235_axes_0, x = var_4806)[name = string("input_235")];
            string input_237_pad_type_0 = const()[name = string("input_237_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_237_strides_0 = const()[name = string("input_237_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_237_pad_0 = const()[name = string("input_237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_237_dilations_0 = const()[name = string("input_237_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_237_groups_0 = const()[name = string("input_237_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_237 = conv(dilations = input_237_dilations_0, groups = input_237_groups_0, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = input_237_strides_0, weight = encoder_layers_23_mlp_gate_proj_weight, x = input_235)[name = string("input_237")];
            string up_47_pad_type_0 = const()[name = string("up_47_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_47_strides_0 = const()[name = string("up_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_47_pad_0 = const()[name = string("up_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_47_dilations_0 = const()[name = string("up_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_47_groups_0 = const()[name = string("up_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_47 = conv(dilations = up_47_dilations_0, groups = up_47_groups_0, pad = up_47_pad_0, pad_type = up_47_pad_type_0, strides = up_47_strides_0, weight = encoder_layers_23_mlp_up_proj_weight, x = input_235)[name = string("up_47")];
            tensor<fp16, [1, 3072, 1, 1024]> var_4820 = silu(x = input_237)[name = string("op_4820")];
            tensor<fp16, [1, 3072, 1, 1024]> input_239 = mul(x = var_4820, y = up_47)[name = string("input_239")];
            string var_4827_pad_type_0 = const()[name = string("op_4827_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4827_strides_0 = const()[name = string("op_4827_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4827_pad_0 = const()[name = string("op_4827_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4827_dilations_0 = const()[name = string("op_4827_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4827_groups_0 = const()[name = string("op_4827_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4827 = conv(dilations = var_4827_dilations_0, groups = var_4827_groups_0, pad = var_4827_pad_0, pad_type = var_4827_pad_type_0, strides = var_4827_strides_0, weight = encoder_layers_23_mlp_down_proj_weight, x = input_239)[name = string("op_4827")];
            tensor<int32, [1]> var_4828_axes_0 = const()[name = string("op_4828_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4828 = squeeze(axes = var_4828_axes_0, x = var_4827)[name = string("op_4828")];
            tensor<int32, [3]> var_4829 = const()[name = string("op_4829"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_47 = transpose(perm = var_4829, x = var_4828)[name = string("transpose_36")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_49_cast_fp16 = add(x = x_809_cast_fp16, y = mlp_out_47)[name = string("hidden_states_49_cast_fp16")];
            fp16 var_5_promoted_96_to_fp16 = const()[name = string("op_5_promoted_96_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_4856_cast_fp16 = pow(x = hidden_states_49_cast_fp16, y = var_5_promoted_96_to_fp16)[name = string("op_4856_cast_fp16")];
            tensor<int32, [1]> var_193_axes_0 = const()[name = string("var_193_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_193_keep_dims_0 = const()[name = string("var_193_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_193_cast_fp16 = reduce_mean(axes = var_193_axes_0, keep_dims = var_193_keep_dims_0, x = var_4856_cast_fp16)[name = string("var_193_cast_fp16")];
            fp16 var_4859_to_fp16 = const()[name = string("op_4859_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_4860_cast_fp16 = add(x = var_193_cast_fp16, y = var_4859_to_fp16)[name = string("op_4860_cast_fp16")];
            fp32 var_4861_epsilon_0 = const()[name = string("op_4861_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_4861_cast_fp16 = rsqrt(epsilon = var_4861_epsilon_0, x = var_4860_cast_fp16)[name = string("op_4861_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_819_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = var_4861_cast_fp16)[name = string("x_819_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_24_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_24_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1192100608)))];
            tensor<fp16, [1, 1024, 1024]> var_4864_cast_fp16 = mul(x = x_819_cast_fp16, y = encoder_layers_24_input_layernorm_weight_promoted_to_fp16)[name = string("op_4864_cast_fp16")];
            tensor<int32, [3]> var_4869 = const()[name = string("op_4869"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_241_axes_0 = const()[name = string("input_241_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4870 = transpose(perm = var_4869, x = var_4864_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 1024, 1, 1024]> input_241 = expand_dims(axes = input_241_axes_0, x = var_4870)[name = string("input_241")];
            string var_4877_pad_type_0 = const()[name = string("op_4877_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4877_strides_0 = const()[name = string("op_4877_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4877_pad_0 = const()[name = string("op_4877_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4877_dilations_0 = const()[name = string("op_4877_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4877_groups_0 = const()[name = string("op_4877_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_4877 = conv(dilations = var_4877_dilations_0, groups = var_4877_groups_0, pad = var_4877_pad_0, pad_type = var_4877_pad_type_0, strides = var_4877_strides_0, weight = encoder_layers_24_self_attn_q_proj_weight, x = input_241)[name = string("op_4877")];
            tensor<int32, [4]> var_4878 = const()[name = string("op_4878"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_4879 = reshape(shape = var_4878, x = var_4877)[name = string("op_4879")];
            tensor<int32, [4]> var_4880 = const()[name = string("op_4880"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4887_pad_type_0 = const()[name = string("op_4887_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4887_strides_0 = const()[name = string("op_4887_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4887_pad_0 = const()[name = string("op_4887_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4887_dilations_0 = const()[name = string("op_4887_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4887_groups_0 = const()[name = string("op_4887_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4887 = conv(dilations = var_4887_dilations_0, groups = var_4887_groups_0, pad = var_4887_pad_0, pad_type = var_4887_pad_type_0, strides = var_4887_strides_0, weight = encoder_layers_24_self_attn_k_proj_weight, x = input_241)[name = string("op_4887")];
            tensor<int32, [4]> var_4888 = const()[name = string("op_4888"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_4889 = reshape(shape = var_4888, x = var_4887)[name = string("op_4889")];
            tensor<int32, [4]> var_4890 = const()[name = string("op_4890"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4897_pad_type_0 = const()[name = string("op_4897_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4897_strides_0 = const()[name = string("op_4897_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4897_pad_0 = const()[name = string("op_4897_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4897_dilations_0 = const()[name = string("op_4897_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4897_groups_0 = const()[name = string("op_4897_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_4897 = conv(dilations = var_4897_dilations_0, groups = var_4897_groups_0, pad = var_4897_pad_0, pad_type = var_4897_pad_type_0, strides = var_4897_strides_0, weight = encoder_layers_24_self_attn_v_proj_weight, x = input_241)[name = string("op_4897")];
            tensor<int32, [4]> var_4898 = const()[name = string("op_4898"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_4899 = reshape(shape = var_4898, x = var_4897)[name = string("op_4899")];
            tensor<int32, [4]> var_4900 = const()[name = string("op_4900"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_97_to_fp16 = const()[name = string("op_5_promoted_97_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_145 = transpose(perm = var_4880, x = var_4879)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 1024, 128]> var_4906_cast_fp16 = pow(x = q_145, y = var_5_promoted_97_to_fp16)[name = string("op_4906_cast_fp16")];
            tensor<int32, [1]> var_195_axes_0 = const()[name = string("var_195_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_195_keep_dims_0 = const()[name = string("var_195_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_195_cast_fp16_0 = reduce_mean(axes = var_195_axes_0, keep_dims = var_195_keep_dims_0, x = var_4906_cast_fp16)[name = string("var_195_cast_fp16")];
            fp16 var_4909_to_fp16 = const()[name = string("op_4909_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_4910_cast_fp16 = add(x = var_195_cast_fp16_0, y = var_4909_to_fp16)[name = string("op_4910_cast_fp16")];
            fp32 var_4911_epsilon_0 = const()[name = string("op_4911_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_4911_cast_fp16 = rsqrt(epsilon = var_4911_epsilon_0, x = var_4910_cast_fp16)[name = string("op_4911_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_827_cast_fp16 = mul(x = q_145, y = var_4911_cast_fp16)[name = string("x_827_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_147 = mul(x = x_827_cast_fp16, y = encoder_layers_24_self_attn_q_norm_weight)[name = string("q_147")];
            fp16 var_5_promoted_98_to_fp16 = const()[name = string("op_5_promoted_98_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_145 = transpose(perm = var_4890, x = var_4889)[name = string("transpose_33")];
            tensor<fp16, [1, 8, 1024, 128]> var_4919_cast_fp16 = pow(x = k_145, y = var_5_promoted_98_to_fp16)[name = string("op_4919_cast_fp16")];
            tensor<int32, [1]> var_197_axes_0 = const()[name = string("var_197_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_197_keep_dims_0 = const()[name = string("var_197_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_197_cast_fp16 = reduce_mean(axes = var_197_axes_0, keep_dims = var_197_keep_dims_0, x = var_4919_cast_fp16)[name = string("var_197_cast_fp16")];
            fp16 var_4922_to_fp16 = const()[name = string("op_4922_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_4923_cast_fp16 = add(x = var_197_cast_fp16, y = var_4922_to_fp16)[name = string("op_4923_cast_fp16")];
            fp32 var_4924_epsilon_0 = const()[name = string("op_4924_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_4924_cast_fp16 = rsqrt(epsilon = var_4924_epsilon_0, x = var_4923_cast_fp16)[name = string("op_4924_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_833_cast_fp16 = mul(x = k_145, y = var_4924_cast_fp16)[name = string("x_833_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_147 = mul(x = x_833_cast_fp16, y = encoder_layers_24_self_attn_k_norm_weight)[name = string("k_147")];
            tensor<fp16, [1, 16, 1024, 128]> var_4928 = mul(x = q_147, y = cos)[name = string("op_4928")];
            tensor<int32, [2]> var_4929_split_sizes_0 = const()[name = string("op_4929_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4929_axis_0 = const()[name = string("op_4929_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_4929_0, tensor<fp16, [1, 16, 1024, 64]> var_4929_1 = split(axis = var_4929_axis_0, split_sizes = var_4929_split_sizes_0, x = q_147)[name = string("op_4929")];
            fp16 const_75_promoted = const()[name = string("const_75_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_4931 = mul(x = var_4929_1, y = const_75_promoted)[name = string("op_4931")];
            bool var_4933_interleave_0 = const()[name = string("op_4933_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_4933 = concat(axis = var_17, interleave = var_4933_interleave_0, values = (var_4931, var_4929_0))[name = string("op_4933")];
            tensor<fp16, [1, 16, 1024, 128]> var_4934 = mul(x = var_4933, y = sin)[name = string("op_4934")];
            tensor<fp16, [1, 16, 1024, 128]> query_49 = add(x = var_4928, y = var_4934)[name = string("query_49")];
            tensor<fp16, [1, 8, 1024, 128]> var_4936 = mul(x = k_147, y = cos)[name = string("op_4936")];
            tensor<int32, [2]> var_4937_split_sizes_0 = const()[name = string("op_4937_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4937_axis_0 = const()[name = string("op_4937_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_4937_0, tensor<fp16, [1, 8, 1024, 64]> var_4937_1 = split(axis = var_4937_axis_0, split_sizes = var_4937_split_sizes_0, x = k_147)[name = string("op_4937")];
            fp16 const_76_promoted = const()[name = string("const_76_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_4939 = mul(x = var_4937_1, y = const_76_promoted)[name = string("op_4939")];
            bool var_4941_interleave_0 = const()[name = string("op_4941_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_4941 = concat(axis = var_17, interleave = var_4941_interleave_0, values = (var_4939, var_4937_0))[name = string("op_4941")];
            tensor<fp16, [1, 8, 1024, 128]> var_4942 = mul(x = var_4941, y = sin)[name = string("op_4942")];
            tensor<fp16, [1, 8, 1024, 128]> x_835 = add(x = var_4936, y = var_4942)[name = string("x_835")];
            tensor<int32, [1]> var_4944_axes_0 = const()[name = string("op_4944_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4944 = expand_dims(axes = var_4944_axes_0, x = x_835)[name = string("op_4944")];
            tensor<int32, [5]> x_837_reps_0 = const()[name = string("x_837_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_837 = tile(reps = x_837_reps_0, x = var_4944)[name = string("x_837")];
            tensor<int32, [4]> var_4947 = const()[name = string("op_4947"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_49 = reshape(shape = var_4947, x = x_837)[name = string("key_49")];
            tensor<int32, [1]> var_4949_axes_0 = const()[name = string("op_4949_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_839 = transpose(perm = var_4900, x = var_4899)[name = string("transpose_32")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4949 = expand_dims(axes = var_4949_axes_0, x = x_839)[name = string("op_4949")];
            tensor<int32, [5]> x_841_reps_0 = const()[name = string("x_841_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_841 = tile(reps = x_841_reps_0, x = var_4949)[name = string("x_841")];
            tensor<int32, [4]> var_4952 = const()[name = string("op_4952"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_49 = reshape(shape = var_4952, x = x_841)[name = string("value_49")];
            bool var_4957_transpose_x_1 = const()[name = string("op_4957_transpose_x_1"), val = bool(false)];
            bool var_4957_transpose_y_1 = const()[name = string("op_4957_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_4957_cast_fp16 = matmul(transpose_x = var_4957_transpose_x_1, transpose_y = var_4957_transpose_y_1, x = query_49, y = key_49)[name = string("op_4957_cast_fp16")];
            fp16 var_4958_to_fp16 = const()[name = string("op_4958_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_145_cast_fp16 = mul(x = var_4957_cast_fp16, y = var_4958_to_fp16)[name = string("attn_weights_145_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_147_cast_fp16 = add(x = attn_weights_145_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_4962_cast_fp16 = softmax(axis = var_17, x = attn_weights_147_cast_fp16)[name = string("op_4962_cast_fp16")];
            bool var_4966_transpose_x_0 = const()[name = string("op_4966_transpose_x_0"), val = bool(false)];
            bool var_4966_transpose_y_0 = const()[name = string("op_4966_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_4966_cast_fp16 = matmul(transpose_x = var_4966_transpose_x_0, transpose_y = var_4966_transpose_y_0, x = var_4962_cast_fp16, y = value_49)[name = string("op_4966_cast_fp16")];
            tensor<int32, [4]> var_4968 = const()[name = string("op_4968"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4971 = const()[name = string("op_4971"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_4969 = transpose(perm = var_4968, x = var_4966_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 1024, 2048]> attn_out_147 = reshape(shape = var_4971, x = var_4969)[name = string("attn_out_147")];
            tensor<int32, [3]> var_4973 = const()[name = string("op_4973"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_24 = const()[name = string("squeeze_24"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1192102720)))];
            string var_4982_pad_type_0 = const()[name = string("op_4982_pad_type_0"), val = string("valid")];
            int32 var_4982_groups_0 = const()[name = string("op_4982_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_4982_strides_0 = const()[name = string("op_4982_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_4982_pad_0 = const()[name = string("op_4982_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_4982_dilations_0 = const()[name = string("op_4982_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_4974 = transpose(perm = var_4973, x = attn_out_147)[name = string("transpose_30")];
            tensor<fp16, [1, 1024, 1024]> var_4982 = conv(dilations = var_4982_dilations_0, groups = var_4982_groups_0, pad = var_4982_pad_0, pad_type = var_4982_pad_type_0, strides = var_4982_strides_0, weight = squeeze_24, x = var_4974)[name = string("op_4982")];
            tensor<int32, [3]> var_4983 = const()[name = string("op_4983"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_149 = transpose(perm = var_4983, x = var_4982)[name = string("transpose_29")];
            tensor<fp16, [1, 1024, 1024]> x_843_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = attn_out_149)[name = string("x_843_cast_fp16")];
            fp16 var_5_promoted_99_to_fp16 = const()[name = string("op_5_promoted_99_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_4989_cast_fp16 = pow(x = x_843_cast_fp16, y = var_5_promoted_99_to_fp16)[name = string("op_4989_cast_fp16")];
            tensor<int32, [1]> var_199_axes_0 = const()[name = string("var_199_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_199_keep_dims_0 = const()[name = string("var_199_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_199_cast_fp16 = reduce_mean(axes = var_199_axes_0, keep_dims = var_199_keep_dims_0, x = var_4989_cast_fp16)[name = string("var_199_cast_fp16")];
            fp16 var_4992_to_fp16 = const()[name = string("op_4992_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_4993_cast_fp16 = add(x = var_199_cast_fp16, y = var_4992_to_fp16)[name = string("op_4993_cast_fp16")];
            fp32 var_4994_epsilon_0 = const()[name = string("op_4994_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_4994_cast_fp16 = rsqrt(epsilon = var_4994_epsilon_0, x = var_4993_cast_fp16)[name = string("op_4994_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_847_cast_fp16 = mul(x = x_843_cast_fp16, y = var_4994_cast_fp16)[name = string("x_847_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_24_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_24_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1196297088)))];
            tensor<fp16, [1, 1024, 1024]> var_4997_cast_fp16 = mul(x = x_847_cast_fp16, y = encoder_layers_24_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4997_cast_fp16")];
            tensor<int32, [3]> var_5002 = const()[name = string("op_5002"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_245_axes_0 = const()[name = string("input_245_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5003 = transpose(perm = var_5002, x = var_4997_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 1024, 1, 1024]> input_245 = expand_dims(axes = input_245_axes_0, x = var_5003)[name = string("input_245")];
            string input_247_pad_type_0 = const()[name = string("input_247_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_247_strides_0 = const()[name = string("input_247_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_247_pad_0 = const()[name = string("input_247_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_247_dilations_0 = const()[name = string("input_247_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_247_groups_0 = const()[name = string("input_247_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_247 = conv(dilations = input_247_dilations_0, groups = input_247_groups_0, pad = input_247_pad_0, pad_type = input_247_pad_type_0, strides = input_247_strides_0, weight = encoder_layers_24_mlp_gate_proj_weight, x = input_245)[name = string("input_247")];
            string up_49_pad_type_0 = const()[name = string("up_49_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_49_strides_0 = const()[name = string("up_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_49_pad_0 = const()[name = string("up_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_49_dilations_0 = const()[name = string("up_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_49_groups_0 = const()[name = string("up_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_49 = conv(dilations = up_49_dilations_0, groups = up_49_groups_0, pad = up_49_pad_0, pad_type = up_49_pad_type_0, strides = up_49_strides_0, weight = encoder_layers_24_mlp_up_proj_weight, x = input_245)[name = string("up_49")];
            tensor<fp16, [1, 3072, 1, 1024]> var_5017 = silu(x = input_247)[name = string("op_5017")];
            tensor<fp16, [1, 3072, 1, 1024]> input_249 = mul(x = var_5017, y = up_49)[name = string("input_249")];
            string var_5024_pad_type_0 = const()[name = string("op_5024_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5024_strides_0 = const()[name = string("op_5024_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5024_pad_0 = const()[name = string("op_5024_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5024_dilations_0 = const()[name = string("op_5024_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5024_groups_0 = const()[name = string("op_5024_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_5024 = conv(dilations = var_5024_dilations_0, groups = var_5024_groups_0, pad = var_5024_pad_0, pad_type = var_5024_pad_type_0, strides = var_5024_strides_0, weight = encoder_layers_24_mlp_down_proj_weight, x = input_249)[name = string("op_5024")];
            tensor<int32, [1]> var_5025_axes_0 = const()[name = string("op_5025_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5025 = squeeze(axes = var_5025_axes_0, x = var_5024)[name = string("op_5025")];
            tensor<int32, [3]> var_5026 = const()[name = string("op_5026"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_49 = transpose(perm = var_5026, x = var_5025)[name = string("transpose_27")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_51_cast_fp16 = add(x = x_843_cast_fp16, y = mlp_out_49)[name = string("hidden_states_51_cast_fp16")];
            fp16 var_5_promoted_100_to_fp16 = const()[name = string("op_5_promoted_100_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_5053_cast_fp16 = pow(x = hidden_states_51_cast_fp16, y = var_5_promoted_100_to_fp16)[name = string("op_5053_cast_fp16")];
            tensor<int32, [1]> var_201_axes_0 = const()[name = string("var_201_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_201_keep_dims_0 = const()[name = string("var_201_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_201_cast_fp16 = reduce_mean(axes = var_201_axes_0, keep_dims = var_201_keep_dims_0, x = var_5053_cast_fp16)[name = string("var_201_cast_fp16")];
            fp16 var_5056_to_fp16 = const()[name = string("op_5056_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_5057_cast_fp16 = add(x = var_201_cast_fp16, y = var_5056_to_fp16)[name = string("op_5057_cast_fp16")];
            fp32 var_5058_epsilon_0 = const()[name = string("op_5058_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_5058_cast_fp16 = rsqrt(epsilon = var_5058_epsilon_0, x = var_5057_cast_fp16)[name = string("op_5058_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_853_cast_fp16 = mul(x = hidden_states_51_cast_fp16, y = var_5058_cast_fp16)[name = string("x_853_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_25_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_25_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1196299200)))];
            tensor<fp16, [1, 1024, 1024]> var_5061_cast_fp16 = mul(x = x_853_cast_fp16, y = encoder_layers_25_input_layernorm_weight_promoted_to_fp16)[name = string("op_5061_cast_fp16")];
            tensor<int32, [3]> var_5066 = const()[name = string("op_5066"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_251_axes_0 = const()[name = string("input_251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5067 = transpose(perm = var_5066, x = var_5061_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 1024, 1, 1024]> input_251 = expand_dims(axes = input_251_axes_0, x = var_5067)[name = string("input_251")];
            string var_5074_pad_type_0 = const()[name = string("op_5074_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5074_strides_0 = const()[name = string("op_5074_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5074_pad_0 = const()[name = string("op_5074_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5074_dilations_0 = const()[name = string("op_5074_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5074_groups_0 = const()[name = string("op_5074_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_5074 = conv(dilations = var_5074_dilations_0, groups = var_5074_groups_0, pad = var_5074_pad_0, pad_type = var_5074_pad_type_0, strides = var_5074_strides_0, weight = encoder_layers_25_self_attn_q_proj_weight, x = input_251)[name = string("op_5074")];
            tensor<int32, [4]> var_5075 = const()[name = string("op_5075"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_5076 = reshape(shape = var_5075, x = var_5074)[name = string("op_5076")];
            tensor<int32, [4]> var_5077 = const()[name = string("op_5077"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5084_pad_type_0 = const()[name = string("op_5084_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5084_strides_0 = const()[name = string("op_5084_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5084_pad_0 = const()[name = string("op_5084_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5084_dilations_0 = const()[name = string("op_5084_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5084_groups_0 = const()[name = string("op_5084_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_5084 = conv(dilations = var_5084_dilations_0, groups = var_5084_groups_0, pad = var_5084_pad_0, pad_type = var_5084_pad_type_0, strides = var_5084_strides_0, weight = encoder_layers_25_self_attn_k_proj_weight, x = input_251)[name = string("op_5084")];
            tensor<int32, [4]> var_5085 = const()[name = string("op_5085"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_5086 = reshape(shape = var_5085, x = var_5084)[name = string("op_5086")];
            tensor<int32, [4]> var_5087 = const()[name = string("op_5087"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5094_pad_type_0 = const()[name = string("op_5094_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5094_strides_0 = const()[name = string("op_5094_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5094_pad_0 = const()[name = string("op_5094_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5094_dilations_0 = const()[name = string("op_5094_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5094_groups_0 = const()[name = string("op_5094_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_5094 = conv(dilations = var_5094_dilations_0, groups = var_5094_groups_0, pad = var_5094_pad_0, pad_type = var_5094_pad_type_0, strides = var_5094_strides_0, weight = encoder_layers_25_self_attn_v_proj_weight, x = input_251)[name = string("op_5094")];
            tensor<int32, [4]> var_5095 = const()[name = string("op_5095"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_5096 = reshape(shape = var_5095, x = var_5094)[name = string("op_5096")];
            tensor<int32, [4]> var_5097 = const()[name = string("op_5097"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_101_to_fp16 = const()[name = string("op_5_promoted_101_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_151 = transpose(perm = var_5077, x = var_5076)[name = string("transpose_25")];
            tensor<fp16, [1, 16, 1024, 128]> var_5103_cast_fp16 = pow(x = q_151, y = var_5_promoted_101_to_fp16)[name = string("op_5103_cast_fp16")];
            tensor<int32, [1]> var_203_axes_0 = const()[name = string("var_203_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_203_keep_dims_0 = const()[name = string("var_203_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_203_cast_fp16 = reduce_mean(axes = var_203_axes_0, keep_dims = var_203_keep_dims_0, x = var_5103_cast_fp16)[name = string("var_203_cast_fp16")];
            fp16 var_5106_to_fp16 = const()[name = string("op_5106_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_5107_cast_fp16 = add(x = var_203_cast_fp16, y = var_5106_to_fp16)[name = string("op_5107_cast_fp16")];
            fp32 var_5108_epsilon_0 = const()[name = string("op_5108_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_5108_cast_fp16 = rsqrt(epsilon = var_5108_epsilon_0, x = var_5107_cast_fp16)[name = string("op_5108_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_861_cast_fp16 = mul(x = q_151, y = var_5108_cast_fp16)[name = string("x_861_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_153 = mul(x = x_861_cast_fp16, y = encoder_layers_25_self_attn_q_norm_weight)[name = string("q_153")];
            fp16 var_5_promoted_102_to_fp16 = const()[name = string("op_5_promoted_102_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_151 = transpose(perm = var_5087, x = var_5086)[name = string("transpose_24")];
            tensor<fp16, [1, 8, 1024, 128]> var_5116_cast_fp16 = pow(x = k_151, y = var_5_promoted_102_to_fp16)[name = string("op_5116_cast_fp16")];
            tensor<int32, [1]> var_205_axes_0 = const()[name = string("var_205_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_205_keep_dims_0 = const()[name = string("var_205_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_205_cast_fp16 = reduce_mean(axes = var_205_axes_0, keep_dims = var_205_keep_dims_0, x = var_5116_cast_fp16)[name = string("var_205_cast_fp16")];
            fp16 var_5119_to_fp16 = const()[name = string("op_5119_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_5120_cast_fp16 = add(x = var_205_cast_fp16, y = var_5119_to_fp16)[name = string("op_5120_cast_fp16")];
            fp32 var_5121_epsilon_0 = const()[name = string("op_5121_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_5121_cast_fp16 = rsqrt(epsilon = var_5121_epsilon_0, x = var_5120_cast_fp16)[name = string("op_5121_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_867_cast_fp16 = mul(x = k_151, y = var_5121_cast_fp16)[name = string("x_867_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_153 = mul(x = x_867_cast_fp16, y = encoder_layers_25_self_attn_k_norm_weight)[name = string("k_153")];
            tensor<fp16, [1, 16, 1024, 128]> var_5125 = mul(x = q_153, y = cos)[name = string("op_5125")];
            tensor<int32, [2]> var_5126_split_sizes_0 = const()[name = string("op_5126_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5126_axis_0 = const()[name = string("op_5126_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_5126_0, tensor<fp16, [1, 16, 1024, 64]> var_5126_1 = split(axis = var_5126_axis_0, split_sizes = var_5126_split_sizes_0, x = q_153)[name = string("op_5126")];
            fp16 const_78_promoted = const()[name = string("const_78_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_5128 = mul(x = var_5126_1, y = const_78_promoted)[name = string("op_5128")];
            bool var_5130_interleave_0 = const()[name = string("op_5130_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_5130 = concat(axis = var_17, interleave = var_5130_interleave_0, values = (var_5128, var_5126_0))[name = string("op_5130")];
            tensor<fp16, [1, 16, 1024, 128]> var_5131 = mul(x = var_5130, y = sin)[name = string("op_5131")];
            tensor<fp16, [1, 16, 1024, 128]> query_51 = add(x = var_5125, y = var_5131)[name = string("query_51")];
            tensor<fp16, [1, 8, 1024, 128]> var_5133 = mul(x = k_153, y = cos)[name = string("op_5133")];
            tensor<int32, [2]> var_5134_split_sizes_0 = const()[name = string("op_5134_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5134_axis_0 = const()[name = string("op_5134_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_5134_0, tensor<fp16, [1, 8, 1024, 64]> var_5134_1 = split(axis = var_5134_axis_0, split_sizes = var_5134_split_sizes_0, x = k_153)[name = string("op_5134")];
            fp16 const_79_promoted = const()[name = string("const_79_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_5136 = mul(x = var_5134_1, y = const_79_promoted)[name = string("op_5136")];
            bool var_5138_interleave_0 = const()[name = string("op_5138_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_5138 = concat(axis = var_17, interleave = var_5138_interleave_0, values = (var_5136, var_5134_0))[name = string("op_5138")];
            tensor<fp16, [1, 8, 1024, 128]> var_5139 = mul(x = var_5138, y = sin)[name = string("op_5139")];
            tensor<fp16, [1, 8, 1024, 128]> x_869 = add(x = var_5133, y = var_5139)[name = string("x_869")];
            tensor<int32, [1]> var_5141_axes_0 = const()[name = string("op_5141_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_5141 = expand_dims(axes = var_5141_axes_0, x = x_869)[name = string("op_5141")];
            tensor<int32, [5]> x_871_reps_0 = const()[name = string("x_871_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_871 = tile(reps = x_871_reps_0, x = var_5141)[name = string("x_871")];
            tensor<int32, [4]> var_5144 = const()[name = string("op_5144"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_51 = reshape(shape = var_5144, x = x_871)[name = string("key_51")];
            tensor<int32, [1]> var_5146_axes_0 = const()[name = string("op_5146_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_873 = transpose(perm = var_5097, x = var_5096)[name = string("transpose_23")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_5146 = expand_dims(axes = var_5146_axes_0, x = x_873)[name = string("op_5146")];
            tensor<int32, [5]> x_875_reps_0 = const()[name = string("x_875_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_875 = tile(reps = x_875_reps_0, x = var_5146)[name = string("x_875")];
            tensor<int32, [4]> var_5149 = const()[name = string("op_5149"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_51 = reshape(shape = var_5149, x = x_875)[name = string("value_51")];
            bool var_5154_transpose_x_1 = const()[name = string("op_5154_transpose_x_1"), val = bool(false)];
            bool var_5154_transpose_y_1 = const()[name = string("op_5154_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_5154_cast_fp16 = matmul(transpose_x = var_5154_transpose_x_1, transpose_y = var_5154_transpose_y_1, x = query_51, y = key_51)[name = string("op_5154_cast_fp16")];
            fp16 var_5155_to_fp16 = const()[name = string("op_5155_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_151_cast_fp16 = mul(x = var_5154_cast_fp16, y = var_5155_to_fp16)[name = string("attn_weights_151_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_153_cast_fp16 = add(x = attn_weights_151_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_153_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_5159_cast_fp16 = softmax(axis = var_17, x = attn_weights_153_cast_fp16)[name = string("op_5159_cast_fp16")];
            bool var_5163_transpose_x_0 = const()[name = string("op_5163_transpose_x_0"), val = bool(false)];
            bool var_5163_transpose_y_0 = const()[name = string("op_5163_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_5163_cast_fp16 = matmul(transpose_x = var_5163_transpose_x_0, transpose_y = var_5163_transpose_y_0, x = var_5159_cast_fp16, y = value_51)[name = string("op_5163_cast_fp16")];
            tensor<int32, [4]> var_5165 = const()[name = string("op_5165"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5168 = const()[name = string("op_5168"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_5166 = transpose(perm = var_5165, x = var_5163_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 1024, 2048]> attn_out_153 = reshape(shape = var_5168, x = var_5166)[name = string("attn_out_153")];
            tensor<int32, [3]> var_5170 = const()[name = string("op_5170"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_25 = const()[name = string("squeeze_25"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1196301312)))];
            string var_5179_pad_type_0 = const()[name = string("op_5179_pad_type_0"), val = string("valid")];
            int32 var_5179_groups_0 = const()[name = string("op_5179_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_5179_strides_0 = const()[name = string("op_5179_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5179_pad_0 = const()[name = string("op_5179_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5179_dilations_0 = const()[name = string("op_5179_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_5171 = transpose(perm = var_5170, x = attn_out_153)[name = string("transpose_21")];
            tensor<fp16, [1, 1024, 1024]> var_5179 = conv(dilations = var_5179_dilations_0, groups = var_5179_groups_0, pad = var_5179_pad_0, pad_type = var_5179_pad_type_0, strides = var_5179_strides_0, weight = squeeze_25, x = var_5171)[name = string("op_5179")];
            tensor<int32, [3]> var_5180 = const()[name = string("op_5180"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_155 = transpose(perm = var_5180, x = var_5179)[name = string("transpose_20")];
            tensor<fp16, [1, 1024, 1024]> x_877_cast_fp16 = add(x = hidden_states_51_cast_fp16, y = attn_out_155)[name = string("x_877_cast_fp16")];
            fp16 var_5_promoted_103_to_fp16 = const()[name = string("op_5_promoted_103_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_5186_cast_fp16 = pow(x = x_877_cast_fp16, y = var_5_promoted_103_to_fp16)[name = string("op_5186_cast_fp16")];
            tensor<int32, [1]> var_207_axes_0 = const()[name = string("var_207_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_207_keep_dims_0 = const()[name = string("var_207_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_207_cast_fp16 = reduce_mean(axes = var_207_axes_0, keep_dims = var_207_keep_dims_0, x = var_5186_cast_fp16)[name = string("var_207_cast_fp16")];
            fp16 var_5189_to_fp16 = const()[name = string("op_5189_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_5190_cast_fp16 = add(x = var_207_cast_fp16, y = var_5189_to_fp16)[name = string("op_5190_cast_fp16")];
            fp32 var_5191_epsilon_0 = const()[name = string("op_5191_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_5191_cast_fp16 = rsqrt(epsilon = var_5191_epsilon_0, x = var_5190_cast_fp16)[name = string("op_5191_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_881_cast_fp16 = mul(x = x_877_cast_fp16, y = var_5191_cast_fp16)[name = string("x_881_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_25_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_25_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1200495680)))];
            tensor<fp16, [1, 1024, 1024]> var_5194_cast_fp16 = mul(x = x_881_cast_fp16, y = encoder_layers_25_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_5194_cast_fp16")];
            tensor<int32, [3]> var_5199 = const()[name = string("op_5199"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_255_axes_0 = const()[name = string("input_255_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5200 = transpose(perm = var_5199, x = var_5194_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 1024, 1, 1024]> input_255 = expand_dims(axes = input_255_axes_0, x = var_5200)[name = string("input_255")];
            string input_257_pad_type_0 = const()[name = string("input_257_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_257_strides_0 = const()[name = string("input_257_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_257_pad_0 = const()[name = string("input_257_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_257_dilations_0 = const()[name = string("input_257_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_257_groups_0 = const()[name = string("input_257_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_257 = conv(dilations = input_257_dilations_0, groups = input_257_groups_0, pad = input_257_pad_0, pad_type = input_257_pad_type_0, strides = input_257_strides_0, weight = encoder_layers_25_mlp_gate_proj_weight, x = input_255)[name = string("input_257")];
            string up_51_pad_type_0 = const()[name = string("up_51_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_51_strides_0 = const()[name = string("up_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_51_pad_0 = const()[name = string("up_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_51_dilations_0 = const()[name = string("up_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_51_groups_0 = const()[name = string("up_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_51 = conv(dilations = up_51_dilations_0, groups = up_51_groups_0, pad = up_51_pad_0, pad_type = up_51_pad_type_0, strides = up_51_strides_0, weight = encoder_layers_25_mlp_up_proj_weight, x = input_255)[name = string("up_51")];
            tensor<fp16, [1, 3072, 1, 1024]> var_5214 = silu(x = input_257)[name = string("op_5214")];
            tensor<fp16, [1, 3072, 1, 1024]> input_259 = mul(x = var_5214, y = up_51)[name = string("input_259")];
            string var_5221_pad_type_0 = const()[name = string("op_5221_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5221_strides_0 = const()[name = string("op_5221_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5221_pad_0 = const()[name = string("op_5221_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5221_dilations_0 = const()[name = string("op_5221_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5221_groups_0 = const()[name = string("op_5221_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_5221 = conv(dilations = var_5221_dilations_0, groups = var_5221_groups_0, pad = var_5221_pad_0, pad_type = var_5221_pad_type_0, strides = var_5221_strides_0, weight = encoder_layers_25_mlp_down_proj_weight, x = input_259)[name = string("op_5221")];
            tensor<int32, [1]> var_5222_axes_0 = const()[name = string("op_5222_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5222 = squeeze(axes = var_5222_axes_0, x = var_5221)[name = string("op_5222")];
            tensor<int32, [3]> var_5223 = const()[name = string("op_5223"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_51 = transpose(perm = var_5223, x = var_5222)[name = string("transpose_18")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_53_cast_fp16 = add(x = x_877_cast_fp16, y = mlp_out_51)[name = string("hidden_states_53_cast_fp16")];
            fp16 var_5_promoted_104_to_fp16 = const()[name = string("op_5_promoted_104_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_5250_cast_fp16 = pow(x = hidden_states_53_cast_fp16, y = var_5_promoted_104_to_fp16)[name = string("op_5250_cast_fp16")];
            tensor<int32, [1]> var_209_axes_0 = const()[name = string("var_209_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_209_keep_dims_0 = const()[name = string("var_209_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_209_cast_fp16 = reduce_mean(axes = var_209_axes_0, keep_dims = var_209_keep_dims_0, x = var_5250_cast_fp16)[name = string("var_209_cast_fp16")];
            fp16 var_5253_to_fp16 = const()[name = string("op_5253_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_5254_cast_fp16 = add(x = var_209_cast_fp16, y = var_5253_to_fp16)[name = string("op_5254_cast_fp16")];
            fp32 var_5255_epsilon_0 = const()[name = string("op_5255_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_5255_cast_fp16 = rsqrt(epsilon = var_5255_epsilon_0, x = var_5254_cast_fp16)[name = string("op_5255_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_887_cast_fp16 = mul(x = hidden_states_53_cast_fp16, y = var_5255_cast_fp16)[name = string("x_887_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_26_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_26_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1200497792)))];
            tensor<fp16, [1, 1024, 1024]> var_5258_cast_fp16 = mul(x = x_887_cast_fp16, y = encoder_layers_26_input_layernorm_weight_promoted_to_fp16)[name = string("op_5258_cast_fp16")];
            tensor<int32, [3]> var_5263 = const()[name = string("op_5263"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_261_axes_0 = const()[name = string("input_261_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5264 = transpose(perm = var_5263, x = var_5258_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 1024, 1, 1024]> input_261 = expand_dims(axes = input_261_axes_0, x = var_5264)[name = string("input_261")];
            string var_5271_pad_type_0 = const()[name = string("op_5271_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5271_strides_0 = const()[name = string("op_5271_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5271_pad_0 = const()[name = string("op_5271_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5271_dilations_0 = const()[name = string("op_5271_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5271_groups_0 = const()[name = string("op_5271_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_5271 = conv(dilations = var_5271_dilations_0, groups = var_5271_groups_0, pad = var_5271_pad_0, pad_type = var_5271_pad_type_0, strides = var_5271_strides_0, weight = encoder_layers_26_self_attn_q_proj_weight, x = input_261)[name = string("op_5271")];
            tensor<int32, [4]> var_5272 = const()[name = string("op_5272"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_5273 = reshape(shape = var_5272, x = var_5271)[name = string("op_5273")];
            tensor<int32, [4]> var_5274 = const()[name = string("op_5274"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5281_pad_type_0 = const()[name = string("op_5281_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5281_strides_0 = const()[name = string("op_5281_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5281_pad_0 = const()[name = string("op_5281_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5281_dilations_0 = const()[name = string("op_5281_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5281_groups_0 = const()[name = string("op_5281_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_5281 = conv(dilations = var_5281_dilations_0, groups = var_5281_groups_0, pad = var_5281_pad_0, pad_type = var_5281_pad_type_0, strides = var_5281_strides_0, weight = encoder_layers_26_self_attn_k_proj_weight, x = input_261)[name = string("op_5281")];
            tensor<int32, [4]> var_5282 = const()[name = string("op_5282"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_5283 = reshape(shape = var_5282, x = var_5281)[name = string("op_5283")];
            tensor<int32, [4]> var_5284 = const()[name = string("op_5284"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5291_pad_type_0 = const()[name = string("op_5291_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5291_strides_0 = const()[name = string("op_5291_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5291_pad_0 = const()[name = string("op_5291_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5291_dilations_0 = const()[name = string("op_5291_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5291_groups_0 = const()[name = string("op_5291_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_5291 = conv(dilations = var_5291_dilations_0, groups = var_5291_groups_0, pad = var_5291_pad_0, pad_type = var_5291_pad_type_0, strides = var_5291_strides_0, weight = encoder_layers_26_self_attn_v_proj_weight, x = input_261)[name = string("op_5291")];
            tensor<int32, [4]> var_5292 = const()[name = string("op_5292"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_5293 = reshape(shape = var_5292, x = var_5291)[name = string("op_5293")];
            tensor<int32, [4]> var_5294 = const()[name = string("op_5294"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_105_to_fp16 = const()[name = string("op_5_promoted_105_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_157 = transpose(perm = var_5274, x = var_5273)[name = string("transpose_16")];
            tensor<fp16, [1, 16, 1024, 128]> var_5300_cast_fp16 = pow(x = q_157, y = var_5_promoted_105_to_fp16)[name = string("op_5300_cast_fp16")];
            tensor<int32, [1]> var_211_axes_0 = const()[name = string("var_211_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_211_keep_dims_0 = const()[name = string("var_211_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_211_cast_fp16 = reduce_mean(axes = var_211_axes_0, keep_dims = var_211_keep_dims_0, x = var_5300_cast_fp16)[name = string("var_211_cast_fp16")];
            fp16 var_5303_to_fp16 = const()[name = string("op_5303_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_5304_cast_fp16 = add(x = var_211_cast_fp16, y = var_5303_to_fp16)[name = string("op_5304_cast_fp16")];
            fp32 var_5305_epsilon_0 = const()[name = string("op_5305_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_5305_cast_fp16 = rsqrt(epsilon = var_5305_epsilon_0, x = var_5304_cast_fp16)[name = string("op_5305_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_895_cast_fp16 = mul(x = q_157, y = var_5305_cast_fp16)[name = string("x_895_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_159 = mul(x = x_895_cast_fp16, y = encoder_layers_26_self_attn_q_norm_weight)[name = string("q_159")];
            fp16 var_5_promoted_106_to_fp16 = const()[name = string("op_5_promoted_106_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_157 = transpose(perm = var_5284, x = var_5283)[name = string("transpose_15")];
            tensor<fp16, [1, 8, 1024, 128]> var_5313_cast_fp16 = pow(x = k_157, y = var_5_promoted_106_to_fp16)[name = string("op_5313_cast_fp16")];
            tensor<int32, [1]> var_213_axes_0 = const()[name = string("var_213_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_213_keep_dims_0 = const()[name = string("var_213_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_213_cast_fp16 = reduce_mean(axes = var_213_axes_0, keep_dims = var_213_keep_dims_0, x = var_5313_cast_fp16)[name = string("var_213_cast_fp16")];
            fp16 var_5316_to_fp16 = const()[name = string("op_5316_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_5317_cast_fp16 = add(x = var_213_cast_fp16, y = var_5316_to_fp16)[name = string("op_5317_cast_fp16")];
            fp32 var_5318_epsilon_0 = const()[name = string("op_5318_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_5318_cast_fp16 = rsqrt(epsilon = var_5318_epsilon_0, x = var_5317_cast_fp16)[name = string("op_5318_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_901_cast_fp16 = mul(x = k_157, y = var_5318_cast_fp16)[name = string("x_901_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_159 = mul(x = x_901_cast_fp16, y = encoder_layers_26_self_attn_k_norm_weight)[name = string("k_159")];
            tensor<fp16, [1, 16, 1024, 128]> var_5322 = mul(x = q_159, y = cos)[name = string("op_5322")];
            tensor<int32, [2]> var_5323_split_sizes_0 = const()[name = string("op_5323_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5323_axis_0 = const()[name = string("op_5323_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_5323_0, tensor<fp16, [1, 16, 1024, 64]> var_5323_1 = split(axis = var_5323_axis_0, split_sizes = var_5323_split_sizes_0, x = q_159)[name = string("op_5323")];
            fp16 const_81_promoted = const()[name = string("const_81_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_5325 = mul(x = var_5323_1, y = const_81_promoted)[name = string("op_5325")];
            bool var_5327_interleave_0 = const()[name = string("op_5327_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_5327 = concat(axis = var_17, interleave = var_5327_interleave_0, values = (var_5325, var_5323_0))[name = string("op_5327")];
            tensor<fp16, [1, 16, 1024, 128]> var_5328 = mul(x = var_5327, y = sin)[name = string("op_5328")];
            tensor<fp16, [1, 16, 1024, 128]> query_53 = add(x = var_5322, y = var_5328)[name = string("query_53")];
            tensor<fp16, [1, 8, 1024, 128]> var_5330 = mul(x = k_159, y = cos)[name = string("op_5330")];
            tensor<int32, [2]> var_5331_split_sizes_0 = const()[name = string("op_5331_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5331_axis_0 = const()[name = string("op_5331_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_5331_0, tensor<fp16, [1, 8, 1024, 64]> var_5331_1 = split(axis = var_5331_axis_0, split_sizes = var_5331_split_sizes_0, x = k_159)[name = string("op_5331")];
            fp16 const_82_promoted = const()[name = string("const_82_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_5333 = mul(x = var_5331_1, y = const_82_promoted)[name = string("op_5333")];
            bool var_5335_interleave_0 = const()[name = string("op_5335_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_5335 = concat(axis = var_17, interleave = var_5335_interleave_0, values = (var_5333, var_5331_0))[name = string("op_5335")];
            tensor<fp16, [1, 8, 1024, 128]> var_5336 = mul(x = var_5335, y = sin)[name = string("op_5336")];
            tensor<fp16, [1, 8, 1024, 128]> x_903 = add(x = var_5330, y = var_5336)[name = string("x_903")];
            tensor<int32, [1]> var_5338_axes_0 = const()[name = string("op_5338_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_5338 = expand_dims(axes = var_5338_axes_0, x = x_903)[name = string("op_5338")];
            tensor<int32, [5]> x_905_reps_0 = const()[name = string("x_905_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_905 = tile(reps = x_905_reps_0, x = var_5338)[name = string("x_905")];
            tensor<int32, [4]> var_5341 = const()[name = string("op_5341"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key_53 = reshape(shape = var_5341, x = x_905)[name = string("key_53")];
            tensor<int32, [1]> var_5343_axes_0 = const()[name = string("op_5343_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_907 = transpose(perm = var_5294, x = var_5293)[name = string("transpose_14")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_5343 = expand_dims(axes = var_5343_axes_0, x = x_907)[name = string("op_5343")];
            tensor<int32, [5]> x_909_reps_0 = const()[name = string("x_909_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_909 = tile(reps = x_909_reps_0, x = var_5343)[name = string("x_909")];
            tensor<int32, [4]> var_5346 = const()[name = string("op_5346"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value_53 = reshape(shape = var_5346, x = x_909)[name = string("value_53")];
            bool var_5351_transpose_x_1 = const()[name = string("op_5351_transpose_x_1"), val = bool(false)];
            bool var_5351_transpose_y_1 = const()[name = string("op_5351_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_5351_cast_fp16 = matmul(transpose_x = var_5351_transpose_x_1, transpose_y = var_5351_transpose_y_1, x = query_53, y = key_53)[name = string("op_5351_cast_fp16")];
            fp16 var_5352_to_fp16 = const()[name = string("op_5352_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_157_cast_fp16 = mul(x = var_5351_cast_fp16, y = var_5352_to_fp16)[name = string("attn_weights_157_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_159_cast_fp16 = add(x = attn_weights_157_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_5356_cast_fp16 = softmax(axis = var_17, x = attn_weights_159_cast_fp16)[name = string("op_5356_cast_fp16")];
            bool var_5360_transpose_x_0 = const()[name = string("op_5360_transpose_x_0"), val = bool(false)];
            bool var_5360_transpose_y_0 = const()[name = string("op_5360_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_5360_cast_fp16 = matmul(transpose_x = var_5360_transpose_x_0, transpose_y = var_5360_transpose_y_0, x = var_5356_cast_fp16, y = value_53)[name = string("op_5360_cast_fp16")];
            tensor<int32, [4]> var_5362 = const()[name = string("op_5362"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5365 = const()[name = string("op_5365"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_5363 = transpose(perm = var_5362, x = var_5360_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 1024, 2048]> attn_out_159 = reshape(shape = var_5365, x = var_5363)[name = string("attn_out_159")];
            tensor<int32, [3]> var_5367 = const()[name = string("op_5367"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_26 = const()[name = string("squeeze_26"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1200499904)))];
            string var_5376_pad_type_0 = const()[name = string("op_5376_pad_type_0"), val = string("valid")];
            int32 var_5376_groups_0 = const()[name = string("op_5376_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_5376_strides_0 = const()[name = string("op_5376_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5376_pad_0 = const()[name = string("op_5376_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5376_dilations_0 = const()[name = string("op_5376_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_5368 = transpose(perm = var_5367, x = attn_out_159)[name = string("transpose_12")];
            tensor<fp16, [1, 1024, 1024]> var_5376 = conv(dilations = var_5376_dilations_0, groups = var_5376_groups_0, pad = var_5376_pad_0, pad_type = var_5376_pad_type_0, strides = var_5376_strides_0, weight = squeeze_26, x = var_5368)[name = string("op_5376")];
            tensor<int32, [3]> var_5377 = const()[name = string("op_5377"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out_161 = transpose(perm = var_5377, x = var_5376)[name = string("transpose_11")];
            tensor<fp16, [1, 1024, 1024]> x_911_cast_fp16 = add(x = hidden_states_53_cast_fp16, y = attn_out_161)[name = string("x_911_cast_fp16")];
            fp16 var_5_promoted_107_to_fp16 = const()[name = string("op_5_promoted_107_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_5383_cast_fp16 = pow(x = x_911_cast_fp16, y = var_5_promoted_107_to_fp16)[name = string("op_5383_cast_fp16")];
            tensor<int32, [1]> var_215_axes_0 = const()[name = string("var_215_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_215_keep_dims_0 = const()[name = string("var_215_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_215_cast_fp16 = reduce_mean(axes = var_215_axes_0, keep_dims = var_215_keep_dims_0, x = var_5383_cast_fp16)[name = string("var_215_cast_fp16")];
            fp16 var_5386_to_fp16 = const()[name = string("op_5386_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_5387_cast_fp16 = add(x = var_215_cast_fp16, y = var_5386_to_fp16)[name = string("op_5387_cast_fp16")];
            fp32 var_5388_epsilon_0 = const()[name = string("op_5388_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_5388_cast_fp16 = rsqrt(epsilon = var_5388_epsilon_0, x = var_5387_cast_fp16)[name = string("op_5388_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_915_cast_fp16 = mul(x = x_911_cast_fp16, y = var_5388_cast_fp16)[name = string("x_915_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_26_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_26_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1204694272)))];
            tensor<fp16, [1, 1024, 1024]> var_5391_cast_fp16 = mul(x = x_915_cast_fp16, y = encoder_layers_26_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_5391_cast_fp16")];
            tensor<int32, [3]> var_5396 = const()[name = string("op_5396"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_265_axes_0 = const()[name = string("input_265_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5397 = transpose(perm = var_5396, x = var_5391_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 1024, 1, 1024]> input_265 = expand_dims(axes = input_265_axes_0, x = var_5397)[name = string("input_265")];
            string input_267_pad_type_0 = const()[name = string("input_267_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_267_strides_0 = const()[name = string("input_267_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_267_pad_0 = const()[name = string("input_267_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_267_dilations_0 = const()[name = string("input_267_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_267_groups_0 = const()[name = string("input_267_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_267 = conv(dilations = input_267_dilations_0, groups = input_267_groups_0, pad = input_267_pad_0, pad_type = input_267_pad_type_0, strides = input_267_strides_0, weight = encoder_layers_26_mlp_gate_proj_weight, x = input_265)[name = string("input_267")];
            string up_53_pad_type_0 = const()[name = string("up_53_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_53_strides_0 = const()[name = string("up_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_53_pad_0 = const()[name = string("up_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_53_dilations_0 = const()[name = string("up_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_53_groups_0 = const()[name = string("up_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up_53 = conv(dilations = up_53_dilations_0, groups = up_53_groups_0, pad = up_53_pad_0, pad_type = up_53_pad_type_0, strides = up_53_strides_0, weight = encoder_layers_26_mlp_up_proj_weight, x = input_265)[name = string("up_53")];
            tensor<fp16, [1, 3072, 1, 1024]> var_5411 = silu(x = input_267)[name = string("op_5411")];
            tensor<fp16, [1, 3072, 1, 1024]> input_269 = mul(x = var_5411, y = up_53)[name = string("input_269")];
            string var_5418_pad_type_0 = const()[name = string("op_5418_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5418_strides_0 = const()[name = string("op_5418_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5418_pad_0 = const()[name = string("op_5418_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5418_dilations_0 = const()[name = string("op_5418_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5418_groups_0 = const()[name = string("op_5418_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_5418 = conv(dilations = var_5418_dilations_0, groups = var_5418_groups_0, pad = var_5418_pad_0, pad_type = var_5418_pad_type_0, strides = var_5418_strides_0, weight = encoder_layers_26_mlp_down_proj_weight, x = input_269)[name = string("op_5418")];
            tensor<int32, [1]> var_5419_axes_0 = const()[name = string("op_5419_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5419 = squeeze(axes = var_5419_axes_0, x = var_5418)[name = string("op_5419")];
            tensor<int32, [3]> var_5420 = const()[name = string("op_5420"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out_53 = transpose(perm = var_5420, x = var_5419)[name = string("transpose_9")];
            tensor<fp16, [1, 1024, 1024]> hidden_states_cast_fp16 = add(x = x_911_cast_fp16, y = mlp_out_53)[name = string("hidden_states_cast_fp16")];
            fp16 var_5_promoted_108_to_fp16 = const()[name = string("op_5_promoted_108_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_5447_cast_fp16 = pow(x = hidden_states_cast_fp16, y = var_5_promoted_108_to_fp16)[name = string("op_5447_cast_fp16")];
            tensor<int32, [1]> var_217_axes_0 = const()[name = string("var_217_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_217_keep_dims_0 = const()[name = string("var_217_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_217_cast_fp16 = reduce_mean(axes = var_217_axes_0, keep_dims = var_217_keep_dims_0, x = var_5447_cast_fp16)[name = string("var_217_cast_fp16")];
            fp16 var_5450_to_fp16 = const()[name = string("op_5450_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_5451_cast_fp16 = add(x = var_217_cast_fp16, y = var_5450_to_fp16)[name = string("op_5451_cast_fp16")];
            fp32 var_5452_epsilon_0 = const()[name = string("op_5452_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_5452_cast_fp16 = rsqrt(epsilon = var_5452_epsilon_0, x = var_5451_cast_fp16)[name = string("op_5452_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_921_cast_fp16 = mul(x = hidden_states_cast_fp16, y = var_5452_cast_fp16)[name = string("x_921_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_27_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_27_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1204696384)))];
            tensor<fp16, [1, 1024, 1024]> var_5455_cast_fp16 = mul(x = x_921_cast_fp16, y = encoder_layers_27_input_layernorm_weight_promoted_to_fp16)[name = string("op_5455_cast_fp16")];
            tensor<int32, [3]> var_5460 = const()[name = string("op_5460"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_271_axes_0 = const()[name = string("input_271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5461 = transpose(perm = var_5460, x = var_5455_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 1024, 1, 1024]> input_271 = expand_dims(axes = input_271_axes_0, x = var_5461)[name = string("input_271")];
            string var_5468_pad_type_0 = const()[name = string("op_5468_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5468_strides_0 = const()[name = string("op_5468_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5468_pad_0 = const()[name = string("op_5468_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5468_dilations_0 = const()[name = string("op_5468_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5468_groups_0 = const()[name = string("op_5468_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1024]> var_5468 = conv(dilations = var_5468_dilations_0, groups = var_5468_groups_0, pad = var_5468_pad_0, pad_type = var_5468_pad_type_0, strides = var_5468_strides_0, weight = encoder_layers_27_self_attn_q_proj_weight, x = input_271)[name = string("op_5468")];
            tensor<int32, [4]> var_5469 = const()[name = string("op_5469"), val = tensor<int32, [4]>([1, 16, 128, 1024])];
            tensor<fp16, [1, 16, 128, 1024]> var_5470 = reshape(shape = var_5469, x = var_5468)[name = string("op_5470")];
            tensor<int32, [4]> var_5471 = const()[name = string("op_5471"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5478_pad_type_0 = const()[name = string("op_5478_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5478_strides_0 = const()[name = string("op_5478_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5478_pad_0 = const()[name = string("op_5478_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5478_dilations_0 = const()[name = string("op_5478_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5478_groups_0 = const()[name = string("op_5478_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_5478 = conv(dilations = var_5478_dilations_0, groups = var_5478_groups_0, pad = var_5478_pad_0, pad_type = var_5478_pad_type_0, strides = var_5478_strides_0, weight = encoder_layers_27_self_attn_k_proj_weight, x = input_271)[name = string("op_5478")];
            tensor<int32, [4]> var_5479 = const()[name = string("op_5479"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_5480 = reshape(shape = var_5479, x = var_5478)[name = string("op_5480")];
            tensor<int32, [4]> var_5481 = const()[name = string("op_5481"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5488_pad_type_0 = const()[name = string("op_5488_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5488_strides_0 = const()[name = string("op_5488_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5488_pad_0 = const()[name = string("op_5488_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5488_dilations_0 = const()[name = string("op_5488_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5488_groups_0 = const()[name = string("op_5488_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_5488 = conv(dilations = var_5488_dilations_0, groups = var_5488_groups_0, pad = var_5488_pad_0, pad_type = var_5488_pad_type_0, strides = var_5488_strides_0, weight = encoder_layers_27_self_attn_v_proj_weight, x = input_271)[name = string("op_5488")];
            tensor<int32, [4]> var_5489 = const()[name = string("op_5489"), val = tensor<int32, [4]>([1, 8, 128, 1024])];
            tensor<fp16, [1, 8, 128, 1024]> var_5490 = reshape(shape = var_5489, x = var_5488)[name = string("op_5490")];
            tensor<int32, [4]> var_5491 = const()[name = string("op_5491"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_5_promoted_109_to_fp16 = const()[name = string("op_5_promoted_109_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024, 128]> q_163 = transpose(perm = var_5471, x = var_5470)[name = string("transpose_7")];
            tensor<fp16, [1, 16, 1024, 128]> var_5497_cast_fp16 = pow(x = q_163, y = var_5_promoted_109_to_fp16)[name = string("op_5497_cast_fp16")];
            tensor<int32, [1]> var_219_axes_0 = const()[name = string("var_219_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_219_keep_dims_0 = const()[name = string("var_219_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1]> var_219_cast_fp16 = reduce_mean(axes = var_219_axes_0, keep_dims = var_219_keep_dims_0, x = var_5497_cast_fp16)[name = string("var_219_cast_fp16")];
            fp16 var_5500_to_fp16 = const()[name = string("op_5500_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1024, 1]> var_5501_cast_fp16 = add(x = var_219_cast_fp16, y = var_5500_to_fp16)[name = string("op_5501_cast_fp16")];
            fp32 var_5502_epsilon_0 = const()[name = string("op_5502_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1024, 1]> var_5502_cast_fp16 = rsqrt(epsilon = var_5502_epsilon_0, x = var_5501_cast_fp16)[name = string("op_5502_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> x_929_cast_fp16 = mul(x = q_163, y = var_5502_cast_fp16)[name = string("x_929_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 128]> q_165 = mul(x = x_929_cast_fp16, y = encoder_layers_27_self_attn_q_norm_weight)[name = string("q_165")];
            fp16 var_5_promoted_110_to_fp16 = const()[name = string("op_5_promoted_110_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 1024, 128]> k_163 = transpose(perm = var_5481, x = var_5480)[name = string("transpose_6")];
            tensor<fp16, [1, 8, 1024, 128]> var_5510_cast_fp16 = pow(x = k_163, y = var_5_promoted_110_to_fp16)[name = string("op_5510_cast_fp16")];
            tensor<int32, [1]> var_221_axes_0_0 = const()[name = string("var_221_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_221_keep_dims_0 = const()[name = string("var_221_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 1024, 1]> var_221_cast_fp16 = reduce_mean(axes = var_221_axes_0_0, keep_dims = var_221_keep_dims_0, x = var_5510_cast_fp16)[name = string("var_221_cast_fp16")];
            fp16 var_5513_to_fp16 = const()[name = string("op_5513_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 1024, 1]> var_5514_cast_fp16 = add(x = var_221_cast_fp16, y = var_5513_to_fp16)[name = string("op_5514_cast_fp16")];
            fp32 var_5515_epsilon_0 = const()[name = string("op_5515_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 1024, 1]> var_5515_cast_fp16 = rsqrt(epsilon = var_5515_epsilon_0, x = var_5514_cast_fp16)[name = string("op_5515_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> x_935_cast_fp16 = mul(x = k_163, y = var_5515_cast_fp16)[name = string("x_935_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> k_165 = mul(x = x_935_cast_fp16, y = encoder_layers_27_self_attn_k_norm_weight)[name = string("k_165")];
            tensor<fp16, [1, 16, 1024, 128]> var_5519 = mul(x = q_165, y = cos)[name = string("op_5519")];
            tensor<int32, [2]> var_5520_split_sizes_0 = const()[name = string("op_5520_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5520_axis_0 = const()[name = string("op_5520_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 1024, 64]> var_5520_0, tensor<fp16, [1, 16, 1024, 64]> var_5520_1 = split(axis = var_5520_axis_0, split_sizes = var_5520_split_sizes_0, x = q_165)[name = string("op_5520")];
            fp16 const_84_promoted = const()[name = string("const_84_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 1024, 64]> var_5522 = mul(x = var_5520_1, y = const_84_promoted)[name = string("op_5522")];
            bool var_5524_interleave_0 = const()[name = string("op_5524_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_5524 = concat(axis = var_17, interleave = var_5524_interleave_0, values = (var_5522, var_5520_0))[name = string("op_5524")];
            tensor<fp16, [1, 16, 1024, 128]> var_5525 = mul(x = var_5524, y = sin)[name = string("op_5525")];
            tensor<fp16, [1, 16, 1024, 128]> query = add(x = var_5519, y = var_5525)[name = string("query")];
            tensor<fp16, [1, 8, 1024, 128]> var_5527 = mul(x = k_165, y = cos)[name = string("op_5527")];
            tensor<int32, [2]> var_5528_split_sizes_0 = const()[name = string("op_5528_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5528_axis_0 = const()[name = string("op_5528_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 1024, 64]> var_5528_0, tensor<fp16, [1, 8, 1024, 64]> var_5528_1 = split(axis = var_5528_axis_0, split_sizes = var_5528_split_sizes_0, x = k_165)[name = string("op_5528")];
            fp16 const_85_promoted = const()[name = string("const_85_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 1024, 64]> var_5530 = mul(x = var_5528_1, y = const_85_promoted)[name = string("op_5530")];
            bool var_5532_interleave_0 = const()[name = string("op_5532_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> var_5532 = concat(axis = var_17, interleave = var_5532_interleave_0, values = (var_5530, var_5528_0))[name = string("op_5532")];
            tensor<fp16, [1, 8, 1024, 128]> var_5533 = mul(x = var_5532, y = sin)[name = string("op_5533")];
            tensor<fp16, [1, 8, 1024, 128]> x_937 = add(x = var_5527, y = var_5533)[name = string("x_937")];
            tensor<int32, [1]> var_5535_axes_0 = const()[name = string("op_5535_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_5535 = expand_dims(axes = var_5535_axes_0, x = x_937)[name = string("op_5535")];
            tensor<int32, [5]> x_939_reps_0 = const()[name = string("x_939_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_939 = tile(reps = x_939_reps_0, x = var_5535)[name = string("x_939")];
            tensor<int32, [4]> var_5538 = const()[name = string("op_5538"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> key = reshape(shape = var_5538, x = x_939)[name = string("key")];
            tensor<int32, [1]> var_5540_axes_0 = const()[name = string("op_5540_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1024, 128]> x_941 = transpose(perm = var_5491, x = var_5490)[name = string("transpose_5")];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_5540 = expand_dims(axes = var_5540_axes_0, x = x_941)[name = string("op_5540")];
            tensor<int32, [5]> x_943_reps_0 = const()[name = string("x_943_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> x_943 = tile(reps = x_943_reps_0, x = var_5540)[name = string("x_943")];
            tensor<int32, [4]> var_5543 = const()[name = string("op_5543"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> value = reshape(shape = var_5543, x = x_943)[name = string("value")];
            bool var_5548_transpose_x_1 = const()[name = string("op_5548_transpose_x_1"), val = bool(false)];
            bool var_5548_transpose_y_1 = const()[name = string("op_5548_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 1024, 1024]> var_5548_cast_fp16 = matmul(transpose_x = var_5548_transpose_x_1, transpose_y = var_5548_transpose_y_1, x = query, y = key)[name = string("op_5548_cast_fp16")];
            fp16 var_5549_to_fp16 = const()[name = string("op_5549_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_163_cast_fp16 = mul(x = var_5548_cast_fp16, y = var_5549_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_5553_cast_fp16 = softmax(axis = var_17, x = attn_weights_165_cast_fp16)[name = string("op_5553_cast_fp16")];
            bool var_5557_transpose_x_0 = const()[name = string("op_5557_transpose_x_0"), val = bool(false)];
            bool var_5557_transpose_y_0 = const()[name = string("op_5557_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024, 128]> var_5557_cast_fp16 = matmul(transpose_x = var_5557_transpose_x_0, transpose_y = var_5557_transpose_y_0, x = var_5553_cast_fp16, y = value)[name = string("op_5557_cast_fp16")];
            tensor<int32, [4]> var_5559 = const()[name = string("op_5559"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5562 = const()[name = string("op_5562"), val = tensor<int32, [3]>([1, 1024, 2048])];
            tensor<fp16, [1, 1024, 16, 128]> var_5560 = transpose(perm = var_5559, x = var_5557_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 1024, 2048]> attn_out_165 = reshape(shape = var_5562, x = var_5560)[name = string("attn_out_165")];
            tensor<int32, [3]> var_5564 = const()[name = string("op_5564"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_27 = const()[name = string("squeeze_27"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1204698496)))];
            string var_5573_pad_type_0 = const()[name = string("op_5573_pad_type_0"), val = string("valid")];
            int32 var_5573_groups_0 = const()[name = string("op_5573_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_5573_strides_0 = const()[name = string("op_5573_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5573_pad_0 = const()[name = string("op_5573_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5573_dilations_0 = const()[name = string("op_5573_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 1024]> var_5565 = transpose(perm = var_5564, x = attn_out_165)[name = string("transpose_3")];
            tensor<fp16, [1, 1024, 1024]> var_5573 = conv(dilations = var_5573_dilations_0, groups = var_5573_groups_0, pad = var_5573_pad_0, pad_type = var_5573_pad_type_0, strides = var_5573_strides_0, weight = squeeze_27, x = var_5565)[name = string("op_5573")];
            tensor<int32, [3]> var_5574 = const()[name = string("op_5574"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> attn_out = transpose(perm = var_5574, x = var_5573)[name = string("transpose_2")];
            tensor<fp16, [1, 1024, 1024]> x_945_cast_fp16 = add(x = hidden_states_cast_fp16, y = attn_out)[name = string("x_945_cast_fp16")];
            fp16 var_5_promoted_111_to_fp16 = const()[name = string("op_5_promoted_111_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_5580_cast_fp16 = pow(x = x_945_cast_fp16, y = var_5_promoted_111_to_fp16)[name = string("op_5580_cast_fp16")];
            tensor<int32, [1]> var_223_axes_0 = const()[name = string("var_223_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_223_keep_dims_0 = const()[name = string("var_223_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_223_cast_fp16 = reduce_mean(axes = var_223_axes_0, keep_dims = var_223_keep_dims_0, x = var_5580_cast_fp16)[name = string("var_223_cast_fp16")];
            fp16 var_5583_to_fp16 = const()[name = string("op_5583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_5584_cast_fp16 = add(x = var_223_cast_fp16, y = var_5583_to_fp16)[name = string("op_5584_cast_fp16")];
            fp32 var_5585_epsilon_0 = const()[name = string("op_5585_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_5585_cast_fp16 = rsqrt(epsilon = var_5585_epsilon_0, x = var_5584_cast_fp16)[name = string("op_5585_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_949_cast_fp16 = mul(x = x_945_cast_fp16, y = var_5585_cast_fp16)[name = string("x_949_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_27_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_27_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1208892864)))];
            tensor<fp16, [1, 1024, 1024]> var_5588_cast_fp16 = mul(x = x_949_cast_fp16, y = encoder_layers_27_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_5588_cast_fp16")];
            tensor<int32, [3]> var_5593 = const()[name = string("op_5593"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_275_axes_0 = const()[name = string("input_275_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5594 = transpose(perm = var_5593, x = var_5588_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 1024, 1, 1024]> input_275 = expand_dims(axes = input_275_axes_0, x = var_5594)[name = string("input_275")];
            string input_277_pad_type_0 = const()[name = string("input_277_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_277_strides_0 = const()[name = string("input_277_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_277_pad_0 = const()[name = string("input_277_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_277_dilations_0 = const()[name = string("input_277_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_277_groups_0 = const()[name = string("input_277_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> input_277 = conv(dilations = input_277_dilations_0, groups = input_277_groups_0, pad = input_277_pad_0, pad_type = input_277_pad_type_0, strides = input_277_strides_0, weight = encoder_layers_27_mlp_gate_proj_weight, x = input_275)[name = string("input_277")];
            string up_pad_type_0 = const()[name = string("up_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_strides_0 = const()[name = string("up_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_pad_0 = const()[name = string("up_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_dilations_0 = const()[name = string("up_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_groups_0 = const()[name = string("up_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 1024]> up = conv(dilations = up_dilations_0, groups = up_groups_0, pad = up_pad_0, pad_type = up_pad_type_0, strides = up_strides_0, weight = encoder_layers_27_mlp_up_proj_weight, x = input_275)[name = string("up")];
            tensor<fp16, [1, 3072, 1, 1024]> var_5608 = silu(x = input_277)[name = string("op_5608")];
            tensor<fp16, [1, 3072, 1, 1024]> input = mul(x = var_5608, y = up)[name = string("input")];
            string var_5615_pad_type_0 = const()[name = string("op_5615_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5615_strides_0 = const()[name = string("op_5615_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5615_pad_0 = const()[name = string("op_5615_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5615_dilations_0 = const()[name = string("op_5615_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5615_groups_0 = const()[name = string("op_5615_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 1024]> var_5615 = conv(dilations = var_5615_dilations_0, groups = var_5615_groups_0, pad = var_5615_pad_0, pad_type = var_5615_pad_type_0, strides = var_5615_strides_0, weight = encoder_layers_27_mlp_down_proj_weight, x = input)[name = string("op_5615")];
            tensor<int32, [1]> var_5616_axes_0 = const()[name = string("op_5616_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5616 = squeeze(axes = var_5616_axes_0, x = var_5615)[name = string("op_5616")];
            tensor<int32, [3]> var_5617 = const()[name = string("op_5617"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 1024]> mlp_out = transpose(perm = var_5617, x = var_5616)[name = string("transpose_0")];
            tensor<fp16, [1, 1024, 1024]> x_953_cast_fp16 = add(x = x_945_cast_fp16, y = mlp_out)[name = string("x_953_cast_fp16")];
            fp16 var_5_promoted_112_to_fp16 = const()[name = string("op_5_promoted_112_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 1024]> var_5623_cast_fp16 = pow(x = x_953_cast_fp16, y = var_5_promoted_112_to_fp16)[name = string("op_5623_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_5623_cast_fp16)[name = string("var_cast_fp16")];
            fp16 var_5626_to_fp16 = const()[name = string("op_5626_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_5627_cast_fp16 = add(x = var_cast_fp16, y = var_5626_to_fp16)[name = string("op_5627_cast_fp16")];
            fp32 var_5628_epsilon_0 = const()[name = string("op_5628_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_5628_cast_fp16 = rsqrt(epsilon = var_5628_epsilon_0, x = var_5627_cast_fp16)[name = string("op_5628_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> x_957_cast_fp16 = mul(x = x_953_cast_fp16, y = var_5628_cast_fp16)[name = string("x_957_cast_fp16")];
            tensor<fp16, [1024]> encoder_norm_weight_promoted_to_fp16 = const()[name = string("encoder_norm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1208894976)))];
            tensor<fp16, [1, 1024, 1024]> var_5631_cast_fp16 = mul(x = x_957_cast_fp16, y = encoder_norm_weight_promoted_to_fp16)[name = string("op_5631_cast_fp16")];
            tensor<int32, [1]> mask_axes_0 = const()[name = string("mask_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> mask_cast_fp16 = expand_dims(axes = mask_axes_0, x = attention_mask)[name = string("mask_cast_fp16")];
            tensor<fp16, [1, 1024, 1024]> var_5645_cast_fp16 = mul(x = var_5631_cast_fp16, y = mask_cast_fp16)[name = string("op_5645_cast_fp16")];
            tensor<int32, [1]> summed_axes_0 = const()[name = string("summed_axes_0"), val = tensor<int32, [1]>([1])];
            bool summed_keep_dims_0 = const()[name = string("summed_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 1024]> summed_cast_fp16 = reduce_sum(axes = summed_axes_0, keep_dims = summed_keep_dims_0, x = var_5645_cast_fp16)[name = string("summed_cast_fp16")];
            tensor<int32, [1]> var_5655_axes_0 = const()[name = string("op_5655_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_5655_keep_dims_0 = const()[name = string("op_5655_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 1]> var_5655_cast_fp16 = reduce_sum(axes = var_5655_axes_0, keep_dims = var_5655_keep_dims_0, x = mask_cast_fp16)[name = string("op_5655_cast_fp16")];
            fp16 var_5656_to_fp16 = const()[name = string("op_5656_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 1]> denom_cast_fp16 = maximum(x = var_5655_cast_fp16, y = var_5656_to_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [1, 1024]> pooled_cast_fp16 = real_div(x = summed_cast_fp16, y = denom_cast_fp16)[name = string("pooled_cast_fp16")];
            tensor<fp16, [1, 1024]> var_5659_cast_fp16 = tanh(x = pooled_cast_fp16)[name = string("op_5659_cast_fp16")];
            fp16 var_5660_to_fp16 = const()[name = string("op_5660_to_fp16"), val = fp16(0x1.fcp+6)];
            tensor<fp16, [1, 1024]> var_5661_cast_fp16 = mul(x = var_5659_cast_fp16, y = var_5660_to_fp16)[name = string("op_5661_cast_fp16")];
            tensor<fp16, [1, 1024]> var_5662_cast_fp16 = round(x = var_5661_cast_fp16)[name = string("op_5662_cast_fp16")];
            fp16 var_5663_promoted_to_fp16 = const()[name = string("op_5663_promoted_to_fp16"), val = fp16(-0x1p+7)];
            fp16 var_5664_promoted_to_fp16 = const()[name = string("op_5664_promoted_to_fp16"), val = fp16(0x1.fcp+6)];
            tensor<fp16, [1, 1024]> clip_0_cast_fp16 = clip(alpha = var_5663_promoted_to_fp16, beta = var_5664_promoted_to_fp16, x = var_5662_cast_fp16)[name = string("clip_0_cast_fp16")];
            string var_5670_dtype_0 = const()[name = string("op_5670_dtype_0"), val = string("int8")];
            tensor<int8, [1, 1024]> embedding = cast(dtype = var_5670_dtype_0, x = clip_0_cast_fp16)[name = string("cast_231")];
        } -> (embedding);
}